WO2022001509A1 - 图像优化方法、装置、计算机存储介质以及电子设备 - Google Patents

图像优化方法、装置、计算机存储介质以及电子设备 Download PDF

Info

Publication number
WO2022001509A1
WO2022001509A1 PCT/CN2021/096024 CN2021096024W WO2022001509A1 WO 2022001509 A1 WO2022001509 A1 WO 2022001509A1 CN 2021096024 W CN2021096024 W CN 2021096024W WO 2022001509 A1 WO2022001509 A1 WO 2022001509A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
optimized
target
network
low
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2021/096024
Other languages
English (en)
French (fr)
Inventor
严宇轩
程培
俞刚
傅斌
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Priority to JP2022552468A priority Critical patent/JP7446457B2/ja
Priority to EP21832144.6A priority patent/EP4050511B1/en
Publication of WO2022001509A1 publication Critical patent/WO2022001509A1/zh
Priority to US17/735,948 priority patent/US12175640B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/70—Denoising; Smoothing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/73—Deblurring; Sharpening
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0475—Generative networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/094—Adversarial learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4046—Scaling of whole images or parts thereof, e.g. expanding or contracting using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
    • G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/20—Image enhancement or restoration using local operators
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/60—Image enhancement or restoration using machine learning, e.g. neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/77—Retouching; Inpainting; Scratch removal
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/37—Determination of transform parameters for the alignment of images, i.e. image registration using transform domain methods
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2200/00—Indexing scheme for image data processing or generation, in general
    • G06T2200/24—Indexing scheme for image data processing or generation, in general involving graphical user interfaces [GUIs]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20024—Filtering details
    • G06T2207/20028—Bilateral filtering
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20024—Filtering details
    • G06T2207/20032—Median filtering
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20048—Transform domain processing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30196—Human being; Person
    • G06T2207/30201—Face

Definitions

  • the present application relates to the field of artificial intelligence technology, in particular, to image optimization technology.
  • the image In the process of imaging, transmission and acquisition, the image will inevitably be affected by external interference and imperfect transmission equipment, which will cause more noise in the image, and lose the original details and become blurred. In order to restore the original details of the image, it is necessary to optimize the image.
  • the first method is to repair the noise and blur of the image through one or more image filtering methods.
  • the second method is to optimize through neural networks. It uses neural network to super-score the image to optimize the image, but the above three methods have the problems of poor denoising effect, poor sharpening effect, or adding details that do not conform to the original image into the image, which reduces the user experience.
  • the present application provides an image optimization method, device, computer storage medium, and electronic device, so as to optimize images at least to a certain extent, improve image quality, and further improve user experience.
  • an image optimization method comprising: acquiring an image to be optimized, performing alignment processing on the image to be optimized, to obtain an image to be optimized, and points of each object in a target area of the image to be optimized and aligned Distributed in standard locations; input the aligned images to be optimized into a generation network, and perform feature extraction on the aligned images to be optimized through the generation network to obtain optimized images; wherein the generation network is based on low-quality image pairs. and the joint loss function is obtained by training the generative adversarial deep neural network model to be trained, and the low-quality image pair includes a target image and a low-quality image corresponding to the target image.
  • an image optimization device the device includes: an acquisition module for acquiring an image to be optimized; an alignment module for performing an alignment process on the to-be-optimized image to obtain an alignment to be optimized image, the points of each object in the target area of the image to be optimized and aligned are distributed in standard positions; the optimization module is used to input the aligned image to be optimized into the generation network, and the aligned image to be optimized is processed by the generation network.
  • the generation network is obtained by training a generative adversarial deep neural network model to be trained according to a low-quality image pair and a joint loss function, and the low-quality image pair includes a target image and the target image Corresponding low-quality images.
  • a computer storage medium on which a computer program is stored, and when the computer program is executed by a processor, the image optimization method described in the first aspect above is implemented.
  • an electronic device for image optimization comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the The instructions are executable to perform the image optimization method described in the first aspect.
  • a computer program product for executing the image optimization method described in the first aspect above when the computer program product is executed.
  • FIG. 1 shows a schematic diagram of an exemplary system architecture to which the technical solutions of the embodiments of the present application can be applied;
  • FIG. 2 shows a schematic flowchart of an image optimization method in an exemplary embodiment of the present application
  • FIG. 3 shows a schematic flowchart of obtaining an image to be optimized for alignment according to an embodiment of the present application
  • 4A, 4B, and 4C show schematic interface diagrams of performing alignment processing on a face image according to an embodiment of the present application
  • FIG. 5 shows a schematic flowchart of training a generative adversarial deep neural network model according to an embodiment of the present application
  • FIG. 6 shows a schematic flowchart of acquiring multiple low-quality image pairs according to an embodiment of the present application
  • FIG. 7 shows a schematic structural diagram of a generative adversarial deep neural network model to be trained according to an embodiment of the present application
  • FIG. 8 shows a schematic structural diagram of a generation network according to an embodiment of the present application.
  • FIG. 9 shows a schematic structural diagram of a post-processing network according to an embodiment of the present application.
  • FIG. 10 shows a general flow chart of a training process for generating an adversarial deep neural network model according to a face image to be trained according to an embodiment of the present application
  • Figure 11A, Figure 11B, Figure 11C, Figure 11D, Figure 11E, Figure 11F show interface schematic diagrams of three groups of using the trained generation network to optimize face images according to an embodiment of the present application;
  • FIG. 12 shows a schematic structural diagram of an image optimization apparatus according to an embodiment of the present application.
  • FIGS. 13A, 13B, 13C, and 13D are schematic diagrams of interfaces for optimizing low-quality images according to an embodiment of the present application.
  • FIG. 14 shows a schematic structural diagram of an electronic device according to an embodiment of the present application.
  • FIG. 1 shows a schematic diagram of an exemplary system architecture to which the technical solutions of the embodiments of the present application can be applied.
  • the system architecture 100 may include a mobile terminal 101 , an information transmission terminal 102 , a network 103 and a server 104 .
  • the above-mentioned mobile terminal 101 can be a terminal device with a camera and a display screen such as a mobile phone, a portable computer, a tablet computer, etc.
  • the information transmission terminal 102 can be a kind of intelligent terminal, such as an intelligent electronic device equipped with various operating systems, etc.
  • 103 is a medium used to provide a communication link between the mobile terminal 101 and the server 104, and between the information transmission terminal 102 and the server 104.
  • the network 103 may include various connection types, such as wired communication links, wireless communication links, etc.
  • the network 103 between the mobile terminal 101 and the information transmission terminal 102 may provide a communication link through a wireless network
  • the network 103 between the mobile terminal 101 and the server 104 and the network 103 between the information transmission terminal 102 and the server 104 may be wireless communication links, specifically a mobile network.
  • the numbers of terminals eg, mobile terminals 101, information transmission terminals 102
  • networks and servers in FIG. 1 are merely illustrative. There can be any number of terminals, networks and servers according to implementation needs.
  • the server 104 may be a server cluster composed of multiple servers, etc., and may be used to store information related to image optimization processing.
  • the mobile terminal 101 sends the image to be optimized to the server 104; the server 104 performs alignment processing on the image to be optimized, and obtains the aligned image to be optimized corresponding to the image to be optimized;
  • the server 104 inputs the to-be-optimized aligned image into the generation network, performs feature extraction on the to-be-optimized aligned image through the generation network to obtain an optimized image, and returns the optimized image to the mobile terminal 101 .
  • the generation network is obtained by training the generative adversarial deep neural network model to be trained according to the low-quality image pair and the joint loss function, and the low-quality image pair includes the target image and the low-quality image corresponding to the target image.
  • the mobile terminal 101 after acquiring the image to be optimized, sends the image to be optimized to the information transmission terminal 102; the information transmission terminal 102 aligns the image to be optimized, and obtains the image to be optimized corresponding to the image to be optimized. Align the images, and then send the aligned images to be optimized to the server 104; the server 104 inputs the aligned images to be optimized into the generation network, and performs feature extraction on the aligned images to be optimized through the generation network to obtain the optimized images, and will optimize the images.
  • the generation network is obtained by training the generative adversarial deep neural network model to be trained according to the low-quality image pair and the joint loss function, and the low-quality image pair includes the target image and the low-quality image corresponding to the target image.
  • the mobile terminal 101 after acquiring the image to be optimized, performs alignment processing on the image to be optimized, obtains the image to be optimized corresponding to the image to be optimized, and then sends the image to be optimized to the server 104;
  • the server 104 inputs the to-be-optimized aligned image into the generation network, performs feature extraction on the to-be-optimized aligned image through the generation network to obtain an optimized image, and returns the optimized image to the mobile terminal 101 .
  • the generation network is obtained by training the generative adversarial deep neural network model to be trained according to the low-quality image pair and the joint loss function, and the low-quality image pair includes the target image and the low-quality image corresponding to the target image.
  • the image optimization method provided by the embodiment of the present application is generally executed by the server 104 , and accordingly, the image optimization apparatus is generally set in the server 104 .
  • the terminal may also have functions similar to the server, so as to execute the image optimization solution provided by the embodiments of the present application.
  • the first is to deblur the image through image processing, which requires one or more image filtering methods to The noise and blur of the image are repaired; the second is to use the neural network to clear the image; the third is to use the neural network to perform image super-score for clarity.
  • the above three methods have certain defects correspondingly.
  • For the first method its processing is more one-sided than the neural network. Since it cannot fully fit the noise and fuzzy distribution in reality, it cannot achieve a good denoising effect;
  • For the second method it mainly focuses on the image sharpening methods of general scenes, and the low-quality image methods used are uneven. If the combinations used in the process of low-quality images are not rich enough, the neural network cannot be very good. Fitting the distribution of real blurred images, resulting in poor clarity of the generated images. In addition, for images of different sizes, no normalization is performed, which will also result in different results due to different sizes of regions when processing related images.
  • an image optimization model is determined based on training a generative adversarial deep neural network model, and an image to be processed is optimized through the image optimization model.
  • the low-quality image can be processed by denoising, sharpening, generating details, etc., so that the low-quality image can maintain the characteristics of the original image, and can be clearer, with higher image quality and better user experience.
  • the technical solution of the present application has low cost of processing images and wide application range.
  • the generative adversarial deep neural network model is a type of neural network model. Compared with the traditional neural network model, its main feature is that it has a discriminative network structure in addition to the generative network structure.
  • the generative network is used to generate images, while The discriminative network is used to judge the authenticity of an image (including the target image and the generated image).
  • iterative training is performed by calculating the difference between the generated image and the target image, and the error of the discriminant network judging the image.
  • the network parameters of the generator network are optimized, so that the generated images are close to the target requirements. Therefore, the generative adversarial deep neural network model can generate high-quality images because of the mutual confrontation between the generative network and the discriminant network.
  • the image optimization method provided by the embodiments of the present application is implemented based on a generative confrontation deep neural network model, and relates to the technical field of artificial intelligence.
  • Artificial Intelligence is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
  • artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can respond in a similar way to human intelligence.
  • Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.
  • an image optimization method is provided, which overcomes the defects existing in the related art at least to a certain extent.
  • the execution subject of the image optimization method provided in this embodiment may be a device with computing processing functions, such as a server or a terminal device, or may be executed jointly by a server and a terminal device, wherein the terminal device and the server may be respectively shown in FIG. 1 .
  • Mobile terminal 101 and server 104 are shown.
  • the image optimization method of the present application can be used to optimize any low-quality image, for example, it can process low-quality images of face images, animal images, images of buildings with fixed structures, etc., to restore and improve image details , the image optimization method provided by the embodiments of the present application will be described in detail below, taking the server as the execution subject and taking the low-quality image as an example of a human face image as an example.
  • FIG. 2 shows a schematic flowchart of an image optimization method in an exemplary embodiment of the present application.
  • the image optimization method provided by this embodiment specifically includes the following steps:
  • the image to be optimized is a low-quality image, and the low-quality image is mainly manifested in poor image definition and a lot of noise.
  • the image to be optimized is a low-quality human face image.
  • the low-quality face image may be an image obtained by a user by using a terminal device with a camera and an imaging unit to photograph the face of the target person or the part containing the face, and the face in the image may be presented at any angle , as long as the facial features of the character can be obtained from it; the to-be-optimized image can also be an image containing a human face downloaded by the user through the network, and so on.
  • an alignment process is performed on the to-be-optimized image to obtain an to-be-optimized aligned image.
  • the image to be optimized as a low-quality face image as an example
  • the standard frontal face position ie, the standard position
  • a standard position template can be used to correct the face image, and the standard position template is the point distribution of each object in a specific area, specifically the standard position template of the face image, which is the face area
  • the point coordinates of the facial features when the face is in a standard frontal face position can be obtained according to a large number of face data statistics to form a five-point coordinate template, that is, a standard position template, wherein the five points include two points marking the left and right eyes, and marking the tip of the nose. Point, mark the two points on the left and right corners of the mouth.
  • the average value of all the coordinate information corresponding to the same part can be used as the point coordinates corresponding to the part in the five-point coordinate template, for example, the coordinate information corresponding to the left eye in all face data can be obtained, Then the coordinate information of all the left eyes is added and averaged to obtain the point coordinates of the left eye in the standard position template.
  • the image to be optimized can be aligned according to the standard position template, that is, the face in the image to be optimized is corrected to a standard frontal face position.
  • FIG. 3 shows a schematic flow chart of acquiring an image to be optimized and aligned. As shown in FIG. 3 , the process of acquiring an image to be optimized and aligned includes S301-S303:
  • a target area in the image to be optimized is detected, and the target area is of the same type as the specific area.
  • the image to be optimized needs to be aligned according to the standard position template, it is necessary to determine the target area in the image to be optimized that is of the same type as the specific area corresponding to the standard position template, that is, the specific area It is the same as the object corresponding to the target area, for example, it is a human face area, an animal face area, and so on.
  • the standard position template is the template corresponding to the face area, then the face area needs to be extracted from the image to be optimized, and then aligned according to the standard position template.
  • the image to be optimized may also include other parts of the human body.
  • the half-body photo includes the neck and upper body in addition to the face area.
  • a model that can be used for face recognition can be used for identification, and the face area can be determined by identifying the facial features. Since the face area in the image to be optimized needs to be aligned, the standard position template used is also the template corresponding to the face area.
  • the transformation matrix between the two can be determined according to the image data corresponding to the face region and the data corresponding to the standard position template. , and then, according to the transformation matrix, the face region in the image to be optimized can be modified into a face region aligned with the five-point coordinates in the standard position template.
  • a transformation operation is performed on the image corresponding to the target area according to the transformation matrix, so as to obtain an image to be optimized and aligned.
  • operations such as translation, rotation, scaling, etc. can be performed on the image to be optimized according to the transformation matrix, and then the image to be optimized can be normalized to a shape consistent with the standard frontal face position, to obtain the aligned images to be optimized.
  • the face alignment process can also be performed in reverse, that is, the aligned face can be restored to the original captured face state through the inverse operation of the transformation matrix.
  • Figures 4A, 4B, and 4C show schematic interface diagrams for aligning face images.
  • the image to be optimized is a low-quality face image, in which the face region exists Inclined, not a standard frontal face position;
  • Figure 4B is a standard position template, i.e. five-point coordinates corresponding to the face image; according to the standard position template shown in Figure 4B to the face image to be optimized shown in Figure 4A
  • the face alignment image to be optimized as shown in FIG. 4C can be obtained, in which the face area conforms to the standard frontal face position.
  • the aligned image to be optimized is input into a generation network, and feature extraction is performed on the aligned image to be optimized through the generation network to obtain an optimized image; wherein the generation network is based on low-quality image pairs and The joint loss function is obtained by training the generative adversarial deep neural network model to be trained, and the low-quality image pair includes a target image and a low-quality image corresponding to the target image.
  • the aligned image to be optimized can be input into the generation network, and the feature extraction of the aligned image to be optimized is performed by the generation network to obtain the optimized image, and the optimized image is Denoise and sharpen the image to be optimized to generate an image after facial detail processing.
  • the generation network is a part of the generative adversarial deep neural network model, which can generate an optimized image corresponding to the input aligned image to be optimized. Before using the generative network to generate optimized images, it is necessary to train the generative adversarial deep neural network model to be trained in order to obtain a stable generative network.
  • FIG. 5 shows a schematic flowchart of training a generative adversarial deep neural network model.
  • the training process of training a generative adversarial deep neural network model specifically includes S501-S504:
  • a low-quality image pair can be used as a training sample for training a generative adversarial deep neural network model to be trained, wherein the low-quality image is used as an input sample, and the target image corresponding to the low-quality image is a verification sample , which is used to judge whether the performance of the generated network is stable. That is to say, the low-quality images in each group of low-quality image pairs are images to be optimized, and the target images are images to be optimized.
  • FIG. 6 shows a schematic flowchart of acquiring multiple low-quality image pairs. As shown in FIG. 6 , the process specifically includes S601-S604:
  • a large number of clear images can be acquired as target images in advance, and the specific number can be determined according to actual needs. The more the number, the higher the performance of the model.
  • the clear image is used as the target image.
  • the face regions in each target image can be aligned to obtain aligned images.
  • the face area in the target image can be aligned according to the standard position template. For example, the face area in the target image can be detected first, and then the point coordinates of the facial features in the face area can be extracted. Finally, according to the extracted The point coordinates of the facial features are aligned with those of the facial features in the standard location template to obtain multiple aligned images.
  • a low-quality image pair is formed according to the target image and the low-quality image corresponding to the target image.
  • a low-quality image may be formed by performing low-quality processing on each aligned image.
  • the low-quality processing may include adding noise processing and/or blurring processing, wherein the noise adding processing includes adding one or more of Gaussian noise, Poisson noise, and salt and pepper noise, and the blurring processing includes mean filtering, Gaussian filtering , one or more of median filtering, bilateral filtering, and reducing resolution.
  • the noise types and fuzzification processing methods added in the embodiments of the present application are not limited to the above-mentioned types, and may also include other types of noise and/or fuzzification processing methods, which will not be repeated in this application.
  • ⁇ represents the mean of the distribution
  • ⁇ represents the standard deviation of the distribution
  • ⁇ 2 represents the variance of the distribution.
  • ⁇ and ⁇ can be determined randomly. After the parameters are determined, noise is added to the color value of each pixel in the image according to the probability distribution, and finally the color value of the pixel is scaled to [0, 255] to achieve Gaussian The addition of noise.
  • the parameter ⁇ can be randomly determined. After the parameters are determined, the color value of each pixel in the image can be processed according to the probability distribution of Poisson noise to add Poisson noise.
  • Salt and pepper noise is to randomly add black and white pixels to the image.
  • the number of black and white pixels can be controlled by the signal-to-noise ratio, and the signal-to-noise ratio can be randomly determined.
  • the total number of pixels can be determined according to the signal-to-noise ratio, and then randomly obtain the position of the pixel to add noise in the image area corresponding to the total number of pixels, and set the pixel value of this position to 255 or 0 , and finally repeat the above steps for other pixels in the image area to complete adding salt and pepper noise to the image.
  • the mean filter is to take the average value of the target pixel and surrounding pixels and then fill it to the target pixel. Its expression is shown in formula (3):
  • M represents the size of the coefficient template
  • f(x, y) represents the pixel value of the target pixel in the image and the surrounding pixels corresponding to M
  • s represents all the pixels in the image
  • Gaussian filtering uses the normal distribution to calculate the transformation of each pixel in the image, and its expression is shown in formula (4):
  • u 2 +v 2 represents the blur radius, and its size can be determined randomly. After the blur radius and variance are determined, the color value of each pixel in the image can be converted according to the normal distribution, so as to realize the blurring of the image.
  • Median filtering is to set the pixel value of each pixel to the median of all pixel values in a certain neighborhood window of the point, and the kernel size of the determined domain window can be randomly determined.
  • Bilateral filtering is a different edge-preserving filtering method that considers both the spatial position (spatial kernel) and the pixel value (value field kernel).
  • the kernel size of the spatial position ie the radius of the Gaussian filter
  • the size of the value field kernel can be Determined by random.
  • Reducing the resolution can reduce the image quality by first reducing the resolution randomly and then upsampling back to the original resolution.
  • the degree of reducing the resolution may be determined randomly.
  • a combination of methods for a target image such as a low-quality face image
  • a target image such as a low-quality face image
  • the low-quality images formed by the processed low-quality images are used to train the generative adversarial deep neural network model to be trained, which can improve the processing accuracy of the model for various low-quality images.
  • each low-quality image pair is used as a target image pair respectively, and the low-quality image in the target image pair is input to the generation network in the generative confrontation deep neural network model to be trained to obtain a generated image.
  • the target image in the pair of the generated image and the target image is input to the post-processing network in the generative adversarial deep neural network model to be trained, and the target image in the pair of the generated image and the target image is processed by the post-processing network to construct Joint loss function.
  • parameters of the generative adversarial deep neural network model to be trained are optimized according to the joint loss function to obtain a generative network.
  • FIG. 7 shows a schematic structural diagram of the generative adversarial deep neural network model to be trained.
  • the generative adversarial deep neural network model 700 to be trained includes a generation network 701 and a post-processing network 702, wherein the generation network 701 is used for input
  • the low-quality image is processed to output the generated image
  • the post-processing network 702 is used to construct a joint loss function according to the generated image and the target image output by the generation network 701, and optimize the parameters of the model based on the joint loss function.
  • FIG. 8 shows a schematic structural diagram of a generation network.
  • the generation network 800 provided in this embodiment includes a downsampling layer 801 , a residual network layer 802 and an upsampling layer 803 .
  • the number of residual network layers 802 may be set to multiple, for example, 4, 5, and so on.
  • the downsampling layer 801 may include multiple convolutional layers with different sizes
  • the upsampling layer 803 may also include multiple convolutional layers with different sizes. Since the convolution operation in the downsampling process can be used to extract The deep features of the image, but compared with the input image, multiple convolution operations make the obtained feature map keep getting smaller, resulting in information loss.
  • the size of the feature map can be restored to the size of the input image by up-sampling after the feature extraction at the residual network layer, thereby reducing the loss of original information during the network transmission process, thereby improving the network efficiency.
  • the structure inconsistency or semantic inconsistency in the half output process will ultimately improve the image quality after optimization.
  • the number and size of the convolutional layers included in the down-sampling layer 801 and the up-sampling layer 803 can be set according to actual needs. ⁇ 256 two convolutional layers, then the upsampling layer 803 may include two convolutional layers with sizes of 256 ⁇ 256 and 512 ⁇ 512 sequentially from front to back.
  • Figure 9 shows a schematic structural diagram of the post-processing network.
  • the post-processing network 900 includes a discriminant network 901, a classification network 902 and a segmentation network 903, wherein the discriminant network 901 may include multiple convolutional layers for Feature extraction is performed on the target image and the generated image, and the confidence level of the target image and the generated image is judged;
  • the classification network 902 may be a classification network such as VGG, which is used to perform feature extraction on the target image and the generated image to obtain the corresponding classification.
  • the segmentation network 903 can be a commonly used segmentation network, which is used to segment the target image, and determine the target image and generate the image according to the position information of each object in the target image.
  • the image information corresponding to the position information of the same object in the image for example, segment the face image to obtain the position information of the facial features in the face image, and determine the target image according to the position information of the facial features and generate the image corresponding to the position information of the facial features
  • the image information of the image itself can be ensured by using the segmentation network 903 to ensure the consistency of the pixels of the image itself.
  • a joint loss function can be constructed according to the processing results of the target image and the generated image. Based on the joint loss function, the parameters of the generative adversarial deep neural network model to be trained can be reversely adjusted. After several iterations After training, a generative adversarial deep neural network model with stable loss function and stable performance can be obtained, and then a generative network for optimizing the low-quality images to be optimized can be obtained.
  • a loss function when constructing a joint loss function, can be constructed through a discriminant network, a classification network, and a segmentation network respectively. network, obtain the first discrimination result and the second discrimination result, and construct the first loss function according to the first discrimination result and the second discrimination result; input the target image in the pair of the generated image and the target image to the classification network, and obtain the first image information and the second image information, and construct a second loss function according to the first image information and the second image information; input the target image in the generated image and the target image pair to the segmentation network, and obtain the first partial image information and the second partial image information image information, and construct a third loss function according to the first partial image information and the second partial image information; finally construct a joint loss function according to the first loss function, the second loss function and the third loss function.
  • the generation network G is used to optimize the low-quality image (input image), and output the image after the optimization process as the generated image.
  • the discrimination network D receives the above-mentioned generated image and the target image corresponding to the above-mentioned low-quality image (input image), and discriminates whether an image (including the target image and the generated image) is true or false.
  • the training goal of the discriminant network D is: to discriminate the target image as true, and to discriminate the generated image as false.
  • the training goal of the generation network G is: to optimize the low-quality image (input image) to obtain the generated image that makes the discrimination result of the discriminant network true, that is, the closer the generated image is to the target image, so as to achieve the effect of being fake and real, so the first A loss function includes a discriminator loss function and a generator loss function.
  • the network D is determined based on the generated image to generate a first determination result D (G (z i)) , while the target image to generate a first determination result D (x i), where z i is input
  • the data of the low-quality image to the generation network G(z i ) is the data of the generated image output by the generation network after optimizing the low-quality image
  • xi is the data of the target image corresponding to the low-quality image
  • D(G (z i )) and D(x i ) are the binary classification confidence of the generated image or the target image output by the discriminant network.
  • the discriminator loss function LossD and the generator loss function LossG can be defined as shown in equations (5)-(6), respectively:
  • x i is the data of the target image corresponding to the image of low quality
  • z i is input to the low quality image data generation network
  • i is any one of a low quality image
  • m is the total number of low-quality image pair.
  • the parameters of the generative network can be fixed first, and the parameters of the discriminant network can be optimized according to formula (5), so that the discriminant network has a discriminant accuracy rate of up to Preset threshold; then fix the optimized parameters of the discriminant network unchanged, and optimize the parameters of the generation network according to formula (6), so that the generation network can generate a clear optimized image.
  • the generated image and the target image should be close in terms of low-level pixel values and high-level abstract features. Therefore, in order to ensure that the generated image and the target image are consistent in deep semantics,
  • the generated image and the target image can also be compared through the classification network, and a perceptual loss function can be constructed according to the comparison result, and then the parameters of the classification network and the generation network can be optimized based on the perceptual loss function.
  • the first image information can be obtained by the classification network processing the generated image
  • the second image information can be obtained by processing the target image
  • the second image information can be determined according to the first image information and the second image information corresponding to each low-quality image pair.
  • the loss function, the perceptual loss function can be obtained by the classification network processing the generated image
  • the second image information can be obtained by processing the target image
  • the loss function, the perceptual loss function can be determined according to the first image information and the second image information corresponding to each low-quality image pair.
  • the second loss function is determined according to the first image information and the second image information corresponding to each low-quality image pair. Specifically, the first image information and the second image information corresponding to each low-quality image pair can be subtracted to obtain The image information is poor; and the second loss function is constructed according to the corresponding image information difference of all low-quality image pairs, and the expression of the second loss function is shown in formula (7):
  • V(G(z i )) is the first image information
  • V(x i ) is the second image information
  • i is any low-quality image pair
  • m is the total number of low-quality image pairs.
  • the parameters of the generation network can be optimized based on the second loss function, so that the generated image output by the generation network is close to or the same as the target image.
  • the image information corresponding to the same object in the generated image and the target image can be compared. If the generated image and the target image correspond to the same object The image information is similar or the same, indicating that the generated image is similar or the same as the target image.
  • the target image may be segmented through a segmentation network to obtain position information of each object in the image.
  • the segmentation network can be used to segment the face image to obtain the position information of the facial features segmentation area, including the position information of the left and right eyes, the position information of the nose, and the position information of the mouth; then, according to the position information of the facial features segmentation area
  • the image information of the corresponding area can be determined from the target image and the generated image.
  • the image information corresponding to the position information of each object in the generated image can be used as the first partial image information, and the position information of the target image corresponding to each object can be used.
  • the image information is used as the second partial image information; finally, the L1 norm between the first partial image information and the second partial image information is calculated, that is, the sum of the absolute values of the image information corresponding to the same object is calculated, and based on all low-quality images
  • a third loss function is constructed for the corresponding L1 norm. The expression of the third loss function is shown in formula (8):
  • M is the position information of each object area after segmentation, i is any low-quality image pair, and m is the total number of low-quality image pairs.
  • a plurality of low-quality image pairs can be regarded as a batch of training samples, and according to the training samples, multiple rounds of iterative training can be performed on the generation adversarial deep neural network model to be trained until the preset number of trainings are completed or The individual loss functions tend to converge.
  • the parameters of the model can be optimized through the first loss function, the second loss function and the third loss function in sequence.
  • FIG. 10 shows a general flow chart of the training process of generating an adversarial deep neural network model for training according to the face image, and the target image is the target face image at this time, Low-quality images are low-quality face images. As shown in FIG.
  • the target face image is aligned to obtain the target face alignment image; in S1002, the target face alignment image is subjected to adding noise processing and / or fuzzification to obtain a low-quality face image; in S1003, the low-quality face image is input into the generation network, and processed through the generation network to output a generated face image, and the generated face image is a pair of low-quality face image after optimization; in S1004, the generated face image and the target face image are paired, and input to the discriminant network, classification network and segmentation network in the post-processing network respectively, and through each network Perform feature extraction on the generated face image and the target face image to determine a joint loss function, the joint loss function includes a first loss function, a second loss function and a third loss function corresponding to each network, and further, according to the first loss function.
  • the first loss function, the second loss function and the third loss function optimize the parameters of the generative adversarial deep neural network model to be trained until the generated face image is close to the target face image.
  • the generative network in the generative adversarial deep neural network model to be trained can be used to optimize other face images to be optimized after alignment processing to obtain clear face images without noise and with facial details. , which further improves the user experience.
  • the generation network after obtaining the optimized image output by the generation network, it can be determined whether to perform position reset processing on the optimized image according to the inclination angle of the face in the image to be optimized relative to the position of the standard frontal face, for example, when the optimized image is to be repositioned.
  • the inclination angle of the face in the optimized image relative to the standard frontal face position is small (in terms of visual effect, the difference is not obvious), no processing is required for the optimized image.
  • the face in the optimized image is relative to the standard frontal face
  • the position repositioning process is performed on the image to be optimized. The object is restored to its original position and angle, and finally an optimized image corresponding to the original to-be-optimized image is obtained.
  • FIG. 11A , FIG. 11B , FIG. 11C , FIG. 11D , FIG. 11E , and FIG. 11F show three groups of interface schematic diagrams for optimizing face images by using the trained generation network, as shown in FIG. 11A , as shown in Figure 11C and Figure 11E, the face images to be optimized are displayed. It can be seen that the face images to be optimized have large noise, low definition, and blurred edge contours. After the generated network, the aligned faces to be optimized After aligning the images for processing, a face image with high definition, clear edge outline and rich facial details can be obtained, as shown in the images shown in FIG. 11B , FIG. 11D , and FIG. 11F , respectively.
  • the optimized image (that is, the optimized image) has the characteristics of high definition, including image details, and accurate noise removal on the basis of retaining the characteristics of the original image.
  • the target images used in model training are all high-definition images. Therefore, when optimizing the image to be optimized according to the generated network after training, the edge contour information can be deepened and the missing parts in the image can be supplemented to a certain extent. That is to say, the image optimization method in this application also has certain image completion and restoration capabilities.
  • the image optimization method of the present application is based on the generative adversarial deep neural network model, so it takes less time to optimize the image, and has the characteristics of high scalability and good portability.
  • a processor including a CPU and a GPU
  • the training of the above-mentioned generative adversarial deep neural network model is realized by GPU, or based on the trained generative adversarial deep neural network model, CPU or GPU is used to realize the optimization processing of the image to be optimized, etc.
  • the computer program is executed by the processor, the above-mentioned functions defined by the above-mentioned methods provided in this application are performed.
  • the program can be stored in a computer-readable storage medium, which can be a read-only memory, a magnetic disk, an optical disk, or the like.
  • FIG. 12 shows a schematic structural diagram of an image optimization apparatus in an exemplary embodiment of the present application.
  • the above image optimization apparatus 1200 includes: an acquisition module 1201 , an alignment module 1202 and an optimization module 1203 .
  • the acquisition module 1201 is used to acquire the image to be optimized;
  • the alignment module 1202 is used to perform alignment processing on the image to be optimized to obtain the aligned image to be optimized.
  • the points of each object in the target area of the image to be optimized are distributed in standard position;
  • the optimization module 1203 is used to input the aligned image to be optimized into the generation network, and perform feature extraction on the aligned image to be optimized through the generation network to obtain the optimized image; wherein the generation network is based on
  • the low-quality image pair and the joint loss function are obtained by training the generative adversarial deep neural network model to be trained, and the low-quality image pair includes a target image and a low-quality image corresponding to the target image.
  • the standard position template is the point distribution of each object in a specific area; the alignment module 1201 is configured to: detect a target area in the image to be optimized, and the target area is related to the specific area. The types of the regions are the same; a transformation matrix between the image data of the target region and the standard position template is determined; and a transformation operation is performed on the image of the target region according to the transformation matrix to obtain the to-be-optimized aligned image.
  • the alignment module is configured to: perform alignment processing on the to-be-optimized image according to a standard position template to obtain the to-be-optimized aligned image.
  • the standard location template is the point distribution of each object in a specific area.
  • the alignment module is configured to detect a target area in the to-be-optimized image, where the target area is the same as the target area.
  • the type of the specific area is the same; determine the transformation matrix between the image data of the target area and the standard position template; perform a transformation operation on the image corresponding to the target area according to the transformation matrix to obtain the target area. Optimize aligned images.
  • the image optimization apparatus 1200 further includes: a low-quality image pair acquisition module, configured to acquire a plurality of the low-quality image pairs; a generating image acquisition module, used to obtain each low-quality image pair Respectively as target image pairs, input the low-quality images in the target image pairs to the generation network in the generative adversarial deep neural network model to be trained to obtain the generated images; the loss function building module is used to combine the generated images and all the generated images.
  • the target image in the target image pair is input to the post-processing network in the generative adversarial deep neural network model to be trained, and the generated image and the target image in the target image pair are processed by the post-processing network.
  • the joint loss function is constructed; a model parameter adjustment module is configured to optimize the parameters of the generative adversarial deep neural network model to be trained according to the joint loss function to obtain the generative network.
  • the low-quality image pair acquisition module is configured to: acquire multiple target images, and perform alignment processing on the multiple target images to acquire multiple aligned images; The images are respectively subjected to low-quality processing to obtain a low-quality image corresponding to each of the target images; the low-quality image pair is formed according to the target image and the low-quality image corresponding to the target image.
  • the low-quality processing includes adding noise processing and/or blurring.
  • the adding noise processing includes adding one or more of Gaussian noise, Poisson noise, and salt and pepper noise
  • the blurring processing includes mean filtering, Gaussian filtering, median filtering, and bilateral filtering. , one or more of reducing the resolution.
  • the post-processing network includes a discriminant network, a classification network and a segmentation network;
  • the loss function construction module includes: a first loss function construction unit for combining the generated image and the target image The target image in the center is input to the discrimination network, the first discrimination result and the second discrimination result are obtained, and the first loss function is constructed according to the first discrimination result and the second discrimination result;
  • the second loss function construction unit for inputting the generated image and the target image in the target image pair to the classification network, obtaining first image information and second image information, and according to the first image information and the second image information information constructs a second loss function;
  • a third loss function constructing unit is used to input the generated image and the target image in the target image pair to the segmentation network, and obtain first partial image information and second partial image information , and constructs a third loss function according to the first partial image information and the second partial image information;
  • a joint loss function construction unit is configured to construct a third loss function according to the first loss function, the second loss
  • the second loss function construction unit is configured to: subtract the first image information and the second image information corresponding to each of the low-quality image pairs to obtain the image information difference;
  • the second loss function is constructed from the image information differences corresponding to all the low-quality image pairs.
  • both the target image and the generated image in the target image pair include multiple objects; based on the foregoing solution, the third loss function construction unit is configured to: pass the segmentation network to all the objects.
  • the target image is segmented to obtain the position information of each object in the target image; the image information corresponding to the position information of each object in the generated image is used as the first partial image information, and the The image information corresponding to the position information of each of the objects in the target image is used as the second partial image information.
  • the third loss function construction unit is configured to: calculate the L1 norm between the first partial image information and the second partial image information; The corresponding L1 norm constructs the third loss function.
  • the model parameter adjustment module is configured to: in each round of training, optimize the to-be-to-be-received function through the first loss function, the second loss function, and the third loss function in sequence.
  • the parameters of the generative adversarial deep neural network model are trained to obtain the generative network.
  • the parameter configuration of optimizing the generative adversarial deep neural network model to be trained by the first loss function is: fixing the parameters of the generative network unchanged, according to the first discrimination result and optimizing the parameters of the discriminating network with the second discriminating result; fixing the optimized parameters of the discriminating network unchanged, and optimizing the parameters of the generating network according to the first discriminating result.
  • the generation network includes: a downsampling layer, a residual network layer, and an upsampling layer.
  • an image optimization apparatus may be configured in a terminal device or a server, and when a user requests through the terminal device to perform optimization processing on a selected low-quality image, the image optimization method in the above embodiment may be executed , for an optimized image.
  • FIG. 13A , FIG. 13B , FIG. 13C , and FIG. 13D show schematic diagrams of interfaces for optimizing low-quality images.
  • the user can turn on the camera function in the terminal device, and in the interface "Please aim the camera at the face” can instruct the user to take a picture of the target face; then take a picture of the target face to obtain a low-quality face Image, low-quality face images are displayed in the photo browsing interface, as shown in Figure 13B; the photo browsing interface shown in Figure 13B includes a "reshoot” button and a "optimization processing” button, when the user selects a new image through the "reshoot” button.
  • the user selects the low-quality face image in the photo browsing interface through the "Optimize” button, the user can perform the optimization process on the low-quality face image.
  • the color of the "Optimization Processing” button can be changed, as shown in Figure 13C, to gray; then call the image optimization service to optimize the low-quality face image captured by the user, and obtain The optimized image returns to the optimized photo browsing interface, as shown in Figure 13D.
  • the user may also select an image that has been photographed or downloaded from the atlas for optimization processing, and the specific processing flow is the same as the image optimization flow in the above-mentioned embodiment, which will not be repeated here.
  • FIG. 14 shows a schematic structural diagram of a computer system suitable for implementing the electronic device according to the embodiment of the present application.
  • the computer system 1400 includes a processor 1401, wherein the processor 1401 may include: a graphics processing unit (Graphics Processing Unit, GPU), a central processing unit (Central Processing Unit, CPU), which can be stored in a read-only A program in a memory (Read-Only Memory, ROM) 1402 or a program loaded from a storage section 1408 into a random access memory (Random Access Memory, RAM) 1403 executes various appropriate actions and processes. In the RAM 1403, various programs and data required for system operation are also stored.
  • a processor (GPU/CPU) 1401, a ROM 1402, and a RAM 1403 are connected to each other through a bus 1404.
  • An Input/Output (I/O) interface 1405 is also connected to the bus 1404 .
  • Computer system 1400 also includes input portion 1406 , output portion 1407 , communication portion 1409 , drives 1410 , and removable media 1411 .
  • embodiments of the present application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the method illustrated in the flowchart.
  • the computer program may be downloaded and installed from the network via the communication portion 1409, and/or installed from the removable medium 1411.
  • the computer program is executed by the processor (GPU/CPU) 1401, various functions defined in the system of the present application are executed.
  • the computer system 1400 may further include an AI (Artificial Intelligence, artificial intelligence) processor for processing computing operations related to machine learning.
  • AI Artificial Intelligence, artificial intelligence
  • the computer-readable medium shown in the embodiments of the present application may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two.
  • the computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or a combination of any of the above.
  • Computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Erasable Programmable Read Only Memory (EPROM), flash memory, optical fiber, portable Compact Disc Read-Only Memory (CD-ROM), optical storage device, magnetic storage device, or any suitable of the above The combination.
  • a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code therein.
  • Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • a computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device .
  • Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination of the foregoing.
  • the units involved in the embodiments of the present application may be implemented in software or hardware, and the described units may also be provided in a processor. Among them, the names of these units do not constitute a limitation on the unit itself under certain circumstances.
  • the present application also provides a computer-readable medium.
  • the computer-readable medium may be included in the electronic device described in the above embodiments; it may also exist alone without being assembled into the electronic device. middle.
  • the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by an electronic device, enables the electronic device to implement the methods described in the above-mentioned embodiments.
  • the exemplary embodiments described herein may be implemented by software, or may be implemented by software combined with necessary hardware. Therefore, the technical solutions according to the embodiments of the present application may be embodied in the form of software products, and the software products may be stored in a non-volatile storage medium (which may be CD-ROM, U disk, mobile hard disk, etc.) or on the network , which includes several instructions to cause a computing device (which may be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present application.
  • a computing device which may be a personal computer, a server, a touch terminal, or a network device, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Human Computer Interaction (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

本申请涉及人工智能技术领域，提供了一种图像优化方法、装置、计算机可读存储介质及电子设备。该方法包括：获取待优化图像，对所述待优化图像进行对齐处理，以获取待优化对齐图像，待优化对齐图像的目标区域中各对象的点分布在标准位置；将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；其中所述生成网络是通过采用低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，低质量图像对包括目标图像以及所述目标图像对应的低质量图像。本技术方案能够提高图像优化效率，并能够消除图像中的噪声，生成图像中的细节，得到清晰的优化图像。

Description

图像优化方法、装置、计算机存储介质以及电子设备
本申请要求于2020年6月28日提交中国专利局、申请号202010595618.2、申请名称为“图像优化方法、装置、计算机存储介质以及电子设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域，具体而言，涉及图像优化技术。
背景技术
图像在成像、传输、获取的过程中，不可避免的会受到外界的干扰、传输设备不完善等的影响，使得图像中出现较多的噪声，并且失去原来的细节而变得模糊。为了恢复图像原来的细节，就需要对图像进行优化处理,。
目前，通常采用三种方法对图像进行优化处理，第一种方法是通过一种或多种图像滤波方法对图像的噪声、模糊进行修复，第二种方法是通过神经网络进行优化；第三种是通过神经网络进行图像超分以优化图像，但是上述三种方法存在去噪效果差、清晰化效果不佳或者将不符合原图像的细节添加到图像中的问题，降低了用户体验。
需要说明的是，上述背景技术部分公开的信息仅用于加强对本申请的背景的理解。
发明内容
本申请提供一种图像优化方法、装置、计算机存储介质及电子设备，进而至少在一定程度上对图像进行优化，提高图像质量，进一步提升用户体验。
本申请的其他特性和优点将通过下面的详细描述变得显然，或部分地通过本申请的实践而习得。
根据本申请的一个方面，提供一种图像优化方法，包括：获取待优化图像，对所述待优化图像进行对齐处理，以获取待优化对齐图像，待优化对齐图像的目标区域中各对象的点分布在标准位置；将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；其中所述生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，低质量图像对包括目标图像以及所述目标图像对应的低质量图像。
根据本申请的一个方面，提供了一种图像优化装置，所述装置包括：获取模块，用于获取待优化图像；对齐模块，用于对所述待优化图像进行对齐处理，以获取待优化对齐图像，待优化对齐图像的目标区域中各对象的点分布在标准位置；优化模块，用于将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；其中所述生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，所述低质量图像对包括目标图像以及所述目标图像对应的低质量图像。
根据本申请的一个方面，提供了一种计算机存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现上述第一方面所述的图像优化方法。
根据本申请的一个方面，提供一种用于图像优化的电子设备，包括：处理器；以及存储器，用于存储所述处理器的可执行指令；其中，所述处理器配置为经由执行所述可执行指令来执行上述第一方面所述的图像优化方法。
根据本申请的一个方面，提供一种计算机程序产品，当所述计算机程序产品被执行时，用于执行上述第一方面所述的图像优化方法。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，示出了符合本申请的实施例，并与说明书一起用于解释本申请的原理。显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。在附图中：
图1示出了可以应用本申请实施例的技术方案的示例性系统架构的示意图；
图2示出本申请一示例性实施例中的图像优化方法的流程示意图；
图3示出了根据本申请一实施例的获取待优化对齐图像的流程示意图；
图4A、图4B、图4C示出了根据本申请一实施例的对人脸图像进行对齐处理的界面示意图；
图5示出了根据本申请一实施例的训练生成对抗深度神经网络模型的流程示意图；
图6示出了根据本申请一实施例的获取多个低质量图像对的流程示意图；
图7示出了根据本申请一实施例的待训练生成对抗深度神经网络模型的结构示意图；
图8示出了根据本申请一实施例的生成网络的结构示意图；
图9示出了根据本申请一实施例的后处理网络的结构示意图；
图10示出了根据本申请一实施例的根据人脸图像对待训练生成对抗深度神经网络模型进行训练的流程总图；
图11A、图11B、图11C、图11D、图11E、图11F示出了根据本申请一实施例的三组采用训练后的生成网络对人脸图像进行优化的界面示意图；
图12示出了根据本申请一实施例的图像优化装置的结构示意图；
图13A、图13B、图13C、13D示出了根据本申请一实施例的对低质量图像进行优化的界面示意图；
图14示出了根据本申请一实施例的电子设备的结构示意图。
具体实施方式
图1示出了可以应用本申请实施例的技术方案的示例性系统架构的示意图。
如图1所示，系统架构100可以包括移动终端101、信息传输终端102、网络103以及服务器104。其中，上述移动终端101可以是手机、便携式计算机、平板电脑等具有摄像头和显示屏幕的终端设备；信息传输终端102可以是一种智能终端，如搭载有各种操作系 统的智能电子设备等；网络103用以移动终端101和服务器104之间、信息传输终端102和服务器104之间提供通信链路的介质。网络103可以包括各种连接类型，例如有线通信链路、无线通信链路等等，在本申请实施例中，移动终端101和信息传输终端102之间的网络103可以是通过无线网络提供通信链路；移动终端101和服务器104之间的网络103、信息传输终端102和服务器104之间的网络103可以是无线通信链路，具体地可以是移动网络。
应该理解，图1中的终端(例如移动终端101、信息传输终端102)、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端、网络和服务器。比如服务器104可以是多个服务器组成的服务器集群等，可用于存储与图像优化处理相关的信息。
在本申请的一个实施例中，移动终端101获取待优化图像后，将待优化图像发送至服务器104；服务器104对待优化图像进行对齐处理，获取到与上述待优化图像对应的待优化对齐图像；服务器104将所述待优化对齐图像输入至生成网络中，通过生成网络对待优化对齐图像进行特征提取，以获取优化图像，并将优化图像返回至移动终端101。其中，生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，低质量图像对包括目标图像以及目标图像对应的低质量图像。
在本申请的一个实施例中，移动终端101获取待优化图像后，将待优化图像发送信息传输终端102；信息传输终端102对待优化图像进行对齐处理，获取到与上述待优化图像对应的待优化对齐图像，然后将该待优化对齐图像发送至服务器104；服务器104将所述待优化对齐图像输入至生成网络中，通过生成网络对待优化对齐图像进行特征提取，以获取优化图像，并将优化图像返回至信息传输终端102，进而，信息传输终端102将优化图像返回至移动终端101。其中，生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，低质量图像对包括目标图像以及目标图像对应的低质量图像。
在本申请的一个实施例中，移动终端101获取待优化图像后，对待优化图像进行对齐处理，获取到与上述待优化图像对应的待优化对齐图像，然后将待优化对齐图像发送至服务器104；服务器104将所述待优化对齐图像输入至生成网络中，通过生成网络对待优化对齐图像进行特征提取，以获取优化图像，并将优化图像返回至移动终端101。其中，生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，低质量图像对包括目标图像以及目标图像对应的低质量图像。
需要说明的是，本申请实施例所提供的图像优化方法一般由服务器104执行，相应地，图像优化装置一般设置于服务器104中。但是，在本申请的其它实施例中，终端也可以与服务器具有相似的功能，从而执行本申请实施例所提供的图像优化方案。
在本领域的相关技术中，在对低质量图像进行优化处理时，通常采用三类优化方法，第一种是通过图像处理方式对图像进行去模糊，需要通过一种或几种图像滤波方法对图像的噪声、模糊进行修复；第二种是通过神经网络来进行图像的清晰化；第三种是利用神经网络进行图像超分来进行清晰化。
上述三种方法相应地存在一定缺陷，对于第一种方法，其相较于神经网络的处理较为片面，由于无法完全拟合现实中的噪声、模糊分布，所以无法达到很好的去噪效果；对于第二种方法，主要集中在一般场景类的图像清晰化方法，且使用到的低质量图像方法参差不齐，如果在低质量图像的过程中采用的组合不够丰富的话，神经网络不能很好地拟合真 实模糊图像的分布，导致生成的图像清晰化不佳，另外对于不同大小的图像，没有进行归一化处理，也会导致在处理相关图像时会由于针对区域的大小不一而结果不佳；对于第三种方法，其在进行图像清晰化的同时会给图像带来一些不符合原图像的细节，比如噪声被转化为了清晰的细节，且该方法主要集中在任意图像的超分辨率上，没有针对图像的具体特征进行单独处理。
本申请提供的技术方案中，基于训练生成对抗深度神经网络模型而确定图像优化模型，通过图像优化模型来对待处理图像进行优化。根据本申请提供的技术方案可以通过对低质量图像进行去噪、清晰化、生成细节等处理使得低质量图像既能维持原图像的特征，又可以更清晰，图像质量更高，用户体验更好。同时，本申请的技术方案处理图像的成本低，适用范围广。
其中，生成对抗深度神经网络模型是一种神经网络模型的类型，其相较于传统神经网络模型主要特点是除了拥有生成网络结构外还拥有一个判别网络结构，其中生成网络用来生成图像，而判别网络则用来判断一幅图像(包括目标图像和生成图像)的真假。在训练生成对抗深度神经网络模型的过程中，通过计算生成图像与目标图像的差异，以及判别网络判断图像的误差等进行迭代训练。进而，通过生成网络与判别器网络的对抗式训练过程，优化生成网络的网络参数，从而使得生成图像接近目标需求。从而，生成对抗深度神经网络模型因为其生成网络与判别网络相互对抗的特性，能够生成质量较高的图像。
本申请实施例提供的图像优化方法是基于生成对抗深度神经网络模型实现的，涉及人工智能技术领域。人工智能(Artificial Intelligence,AI)是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。换句话说，人工智能是计算机科学的一个综合技术，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式做出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。
本申请实施例提供的方案涉及人工智能的图像处理技术和图像识别技术，具体通过如下实施例进行说明：
在本申请的实施例中，提供了一种图像优化方法，至少在一定程度上克服相关技术中所存在的缺陷。本实施例提供的图像优化方法的执行主体可以是具有计算处理功能的设备，例如可以是服务器、终端设备，也可以是由服务器和终端设备共同执行，其中终端设备、服务器可以分别为图1中所示的移动终端101和服务器104。本申请的图像优化方法可用于对任意的低质量图像进行优化处理，例如可以对人脸图像、动物图像、具有固定结构的建筑物图像等类型的低质量图像进行处理，以恢复并提高图像细节，下面以服务器为执行主体、以低质量图像是人脸图像为例对本申请实施例提供的图像优化方法进行详细说明。
图2示出本申请一示例性实施例中图像优化方法的流程示意图。参考图2，该实施例提供的图像优化方法具体包括如下步骤：
在S210中，获取待优化图像。
待优化图像为低质量图像，低质量主要表现为图像的清晰度差、噪声多。在本申请的一个实施例中，以低质量图像是人脸图像为例，该待优化图像为低质量人脸图像。该低质量人脸图像可以是用户通过具有摄像头和成像单元的终端设备对目标人物的脸部或包含脸 部的部位进行拍摄所获取的图像，该图像中的人脸可以是以任意角度呈现的，只要能够从中获取人物的五官即可；该待优化图像还可以是用户通过网络下载得到的包含人脸的图像，等等。
在S220中，对所述待优化图像进行对齐处理，以获取待优化对齐图像。
在对待优化图像进行优化前，有必要对待优化图像进行修正，使待优化图像的目标区域中各对象的点分布在标准位置。
在本申请的一个实施例中，继续以待优化图像为低质量人脸图像为例，在对待优化图像进行优化前，有必要对待优化图像中的人脸的角度进行修正，使图像中人脸位于标准正面人脸位置(即标准位置)。
在一种可能的实现方式中，可以采用一个标准位置模板对人脸图像进行修正，该标准位置模板为特定区域中各对象的点分布，具体到人脸图像的标准位置模板，其为人脸区域中五官(即对象)的点分布。例如，可以根据大量的人脸数据统计获取人脸在处于标准正面人脸位置时五官的点坐标形成五点坐标模板，即标准位置模板，其中五点包括标志左右眼的两点、标志鼻尖的点、标志左右嘴角的两点。在计算各个点的坐标时，可以将与同一部位对应的所有坐标信息的平均值作为五点坐标模板中与该部位对应的点坐标，例如可以获取所有人脸数据中对应左眼的坐标信息，然后将所有左眼的坐标信息进行加和平均，以得到标准位置模板中左眼的点坐标。
在本申请的一个实施例中，在确定标准位置模板后，可以根据该标准位置模板对待优化图像进行对齐处理，也就是将待优化图像中的人脸修正为标准正面人脸位置。图3示出了获取待优化对齐图像的流程示意图，如图3所示，获取待优化对齐图像的流程包括S301-S303：
在S301中，检测待优化图像中的目标区域，该目标区域与特定区域的类型相同。
在本申请的一个实施例中，由于需要根据标准位置模板对待优化图像进行对齐处理，因此需要确定待优化图像中与标准位置模板所对应的特定区域类型相同的目标区域，也就是说，特定区域和目标区域所对应的对象相同，例如都是人脸区域、动物的脸部区域等等。举例而言，标准位置模板是对应人脸区域的模板，那么就需要从待优化图像中提取人脸区域，然后再根据标准位置模板对其进行对齐处理。在实际的图像处理过程中，待优化图像中除了包含人脸区域，可能还包括人体其它部位，例如半身照除了包含人脸区域还包含脖子和上半身，为了针对人脸区域进行优化，就需要对待优化图像中的人脸区域进行识别，然后再对识别得到的人脸区域进行优化。在对待优化图像中的人脸区域进行识别时，可以采用可用于人脸识别的模型进行识别，通过对五官进行识别进而确定人脸区域。由于需要对待优化图像中的人脸区域进行对齐，因此采用的标准位置模板也是与人脸区域对应的模板。
在S302中，确定目标区域的图像数据与标准位置模板之间的变换矩阵。
在本申请的一个实施例中，在将待优化图像中的人脸区域与标准位置模板对齐时，可以根据人脸区域对应的图像数据和标准位置模板对应的数据确定二者之间的变换矩阵，进而可以根据变换矩阵将待优化图像中的人脸区域修正为与标准位置模板中的五点坐标对齐的人脸区域。
在S303中，根据变换矩阵对目标区域对应的图像进行变换操作，以获取待优化对齐图像。
在本申请的一个实施例中，在获取变换矩阵后，可以根据变换矩阵对待优化图像进行平移、旋转、缩放等操作，进而将待优化图像归一化到与标准正面人脸位置一致的形态，以获取待优化对齐图像。进一步地，人脸的对齐处理还可以反向进行，也就是说，对齐后的人脸可以通过变换矩阵的逆操作还原到原始拍摄的人脸状态。
图4A、图4B、图4C示出了对人脸图像进行对齐处理的界面示意图，如图4A所示，为待优化图像，该待优化图像为低质量人脸图像，其中的人脸区域存在倾斜，不是标准正面人脸位置；如图4B所示为标准位置模板，即与人脸图像对应的五点坐标；根据图4B所示的标准位置模板对图4A所示的待优化人脸图像进行对齐处理后，能够得到如图4C所示的待优化人脸对齐图像，其中人脸区域符合标准正面人脸位置。
在S230中，将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；其中所述生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，所述低质量图像对包括目标图像以及所述目标图像对应的低质量图像。
在本申请的一个实施例中，在获取待优化对齐图像后，可以将待优化对齐图像输入至生成网络中，通过生成网络对待优化对齐图像进行特征提取，以获取优化图像，该优化图像即为对待优化图像进行去噪、清晰化生成面部细节处理后的图像。
在本申请的一个实施例中，生成网络为生成对抗深度神经网络模型中的一部分，其可根据输入的待优化对齐图像生成与其对应的优化图像。在采用生成网络生成优化图像之前，需要对待训练生成对抗深度神经网络模型进行训练，才能获得稳定的生成网络。
在本申请的一个实施例中，图5示出了训练生成对抗深度神经网络模型的流程示意图，如图5所示，训练生成对抗深度神经网络模型的训练流程具体包括S501-S504：
在S501中，获取多个低质量图像对。
在本申请的一个实施例中，低质量图像对可以作为训练样本，用于对待训练生成对抗深度神经网络模型进行训练，其中低质量图像作为输入样本，与低质量图像对应的目标图像为验证样本，用于判断生成网络的性能是否达到稳定。也就是说，每组低质量图像对中低质量图像是待优化图像，目标图像为期待优化后的图像。
在本申请的一个实施例中，图6示出了获取多个低质量图像对的流程示意图，如图6所示，该流程具体包括S601-S604：
在S601中，获取多个目标图像。
在S602中，对多个目标图像分别进行对齐处理，以获取多个对齐图像。
在本申请的一个实施例中，可以预先获取大量的清晰图像作为目标图像，具体数量可以根据实际需要确定，数量越多，模型的性能越高，例如可以搜集1w、2w等数量的有人脸的清晰图像作为目标图像。在获取目标图像后，可以对各个目标图像中的人脸区域进行对齐处理，以获取对齐图像。在对齐时，可以根据标准位置模板对目标图像中的人脸区域进行对齐处理，例如首先可以对目标图像中的人脸区域进行检测，然后提取人脸区域中五 官的点坐标，最后根据提取到的五官的点坐标和标准位置模板中五官点坐标进行对齐，以获取多个对齐图像。
在S603中，对多个对齐图像分别进行低质量化处理，以获取与每个目标图像分别对应的低质量图像。
在S604中，根据目标图像和与目标图像对应的低质量图像形成低质量图像对。
在本申请的一个实施例中，可以通过对各个对齐图像进行低质量化处理形成低质量图像。低质量化处理例如可以包括增加噪声处理和/或模糊化处理，其中，增加噪声处理包括增加高斯噪声、泊松噪声、椒盐噪声中的一种或多种，模糊化处理包括均值滤波、高斯滤波、中值滤波、双边滤波、降低分辨率中的一种或多种。当然本申请实施例中增加的噪声类型和模糊化处理方式不限于上述类型，还可以包括其它类型的噪音和/或模糊化处理方式，本申请在此不再赘述。
接下来对各类噪声和模糊化处理方法进行简单介绍。
高斯噪声的概率密度服从高斯分布，高斯分布的表达式如公式(1)所示：
Figure PCTCN2021096024-appb-000001
其中，μ代表分布的均值，σ代表分布的标准差，σ 2代表分布的方差。在本申请的实施例中，μ和σ可以随机确定，在确定参数后根据该概率分布对图像中各个像素的颜色值添加噪声，最后将像素的颜色值放缩到[0,255]即可实现高斯噪声的添加。
泊松噪声的概率密度服从泊松分布，泊松分布的表达式如公式(2)所示：
Figure PCTCN2021096024-appb-000002
其中，参数λ可以随机确定。在确定参数后，可以根据泊松噪声的概率分布对图像中各个像素的颜色值进行处理，以添加泊松噪声。
椒盐噪声即随机在图像上添加黑白像素点，黑白像素的多少可以由信噪比来进行控制，信噪比可以随机确定。在指定信噪比后，根据信噪比可以确定总像素数目，然后在与总像素数目对应的图像区域中随机获取要增加噪声的像素的位置，并将该位置的像素值设置为255或0，最后对该图像区域中其它的像素重复上述步骤，即可完成对图像增加椒盐噪声。
均值滤波是对目标像素及周边像素取平均值后再填充至目标像素，其表达式如公式(3)所示：
Figure PCTCN2021096024-appb-000003
其中，M代表系数模板的大小，f(x,y)代表图像中目标像素和对应M的周边像素的像素值，s代表图像中所有的像素，g(x,y)代表对目标像素进行均值滤波后的像素值。举例而 言，若系数模板的大小为3×3，那么M＝9，因此可以对目标像素和其周围的8个像素的像素值求平均值，并将该平均值填充至目标像素，即可实现对图像的模糊化处理。
高斯滤波用正态分布计算图像中每个像素的变换，其表达式如公式(4)所示：
Figure PCTCN2021096024-appb-000004
其中，u 2+v 2代表模糊半径，其大小可以随机确定。在确定模糊半径和方差后，可以根据该正态分布对图像中各个像素的颜色值进行转换，以实现对图像的模糊化处理。
中值滤波是将每一个像素点的像素值设置为该点某邻域窗口内所有像素点像素值的中值，确定领域窗口的核大小可以随机确定。
双边滤波是既考虑空间位置(空域核)不同也考虑像素值(值域核)的不同的边缘保留滤波方法，其中空间位置的核大小(即高斯滤波的半径)以及值域核的大小都可以通过随机确定。
降低分辨率可以先通过随机降低分辨率，然后上采样回到原分辨率以降低图像质量。在本申请的实施例中，降低分辨率的程度可以随机确定。
在本申请的一个实施例中，可以通过以上方法的随机组合确定一个针对目标图像，例如低质量化人脸图像的方法组合，以模拟真实的低质量图像。由于是随机组合，并且每个组合里所涉及到的低质量化方法又存在随机变量，所以每张目标图像被处理后的效果也不尽相同，基于由目标图像和对目标图像进行低质量化处理后的低质量图像所形成的低质量图像对对待训练生成对抗深度神经网络模型进行训练，可以提高模型对各类低质量图像的处理精度。
在S502中，将每个低质量图像对分别作为目标图像对，将目标图像对中的低质量图像输入至待训练生成对抗深度神经网络模型中的生成网络，得到生成图像。
在S503中，将生成图像和目标图像对中的目标图像输入至待训练生成对抗深度神经网络模型中的后处理网络，通过后处理网络对生成图像和目标图像对中的目标图像进行处理以构建联合损失函数。
在S504中，根据联合损失函数优化待训练生成对抗深度神经网络模型的参数，得到生成网络。
在示例性的实施例中，在介绍对待训练生成对抗深度神经网络模型的训练过程之前，先对本实施例采用的待训练生成对抗深度神经网络模型的网络结构进行说明。图7示出了待训练生成对抗深度神经网络模型的结构示意图，如图7所示，待训练生成对抗深度神经网络模型700包括生成网络701和后处理网络702，其中生成网络701用于对输入的低质量图像进行处理，以输出生成图像，后处理网络702用于根据生成网络701输出的生成图像和目标图像构建联合损失函数，并基于联合损失函数优化模型的参数。
图8示出了生成网络的结构示意图，如图8所示，该实施例提供的生成网络800包括下采样层801、残差网络层802和上采样层803。其中，残差网络层802的数量可以设置为多个，例如4个、5个，等等。在本申请的实施例中，下采样层801可以包括多个大小不 同的卷积层，上采样层803也包括多个大小不同的卷积层，由于下采样过程中的卷积操作可用于提取图像的深层特征，但是相较于输入图像，多次的卷积操作使得得到的特征图不断变小，造成信息丢失。因此为了减少信息的丢失，可以在残差网络层进行特征提取后通过上采样的方式将特征图的大小恢复到输入图像的大小，进而能够减少网络传递过程中原始信息的丢失，从而改善网络后半部分输出过程中的结构不一致或者语义不一致的现象，最终提高优化处理后的图像质量。在本申请的实施例中，下采样层801和上采样层803中包含的卷积层的数量和大小可以根据实际需要设置，例如下采样层801从前至后依次包括大小为512×512、256×256两层卷积层，那么上采样层803可以从前至后依次包括大小为256×256、512×512两层卷积层。
图9示出了后处理网络的结构示意图，如图9所示，后处理网络900包括判别网络901、分类网络902和分割网络903，其中判别网络901可以包括多个卷积层，用于对目标图像和生成图像进行特征提取，并判断目标图像和生成图像的置信度；分类网络902具体可以是VGG等用于分类的网络，用于对目标图像和生成图像进行特征提取以获取对应的分类结果，通过采用分类网络902能够保证图像深层语义的一致性；分割网络903可以是常用的分割网络，用于对目标图像进行分割，并根据目标图像中各个对象的位置信息确定目标图像和生成图像中与同一对象的位置信息对应的图像信息，例如对人脸图像进行分割，以获取人脸图像中五官的位置信息，并根据五官的位置信息确定目标图像和生成图像中与五官的位置信息对应的图像信息，通过采用分割网络903能够保证图像本身像素的一致性。通过判别网络901、分类网络902和分割网络903的相互作用，能够使得生成图像更接近目标图像，提高生成网络的性能。
在本申请的一个实施例中，根据对目标图像和生成图像的处理结果可以构建联合损失函数，基于该联合损失函数可以对待训练生成对抗深度神经网络模型进行反向调参，在经过多次迭代训练后，即可得到损失函数收敛，性能稳定的生成对抗深度神经网络模型，进而可以获取用于对低质量的待优化图像进行优化处理的生成网络。
在本申请的一个实施例中，在构建联合损失函数时，可以通过判别网络、分类网络和分割网络分别构建一个损失函数，具体地，可以将生成图像和目标图像对中的目标图像输入至判别网络，获取第一判别结果和第二判别结果，并根据第一判别结果和第二判别结果构建第一损失函数；将生成图像和目标图像对中的目标图像输入至分类网络，获取第一图像信息和第二图像信息，并根据第一图像信息和第二图像信息构建第二损失函数；将生成图像和目标图像对中的目标图像输入至分割网络，获取第一局部图像信息和第二局部图像信息，并根据第一局部图像信息和第二局部图像信息构建第三损失函数；最后根据第一损失函数、第二损失函数和第三损失函数构建联合损失函数。
在本申请的一个实施例中，生成网络G用来对低质量图像(输入图像)进行优化处理，并将其优化处理之后的图像输出作为生成图像。而判别网络D接收上述生成图像以及与上述低质量图像(输入图像)对应的目标图像，并判别一幅图像(包括目标图像和生成图像)的真假。同时，判别网络D的训练目标是：判别目标图像为真，判别生成图像为假。而生成网络G的训练目标是：对于低质量图像(输入图像)进行优化处理得到令判别网络的判别结果为真的生成图像，即使得生成图像越接近目标图像，以达到以假乱真的效果，因此第一损失函数包括判别器损失函数和生成器损失函数。
在本申请的一个实施例中，判别网络D根据生成图像生成第一判别结果D(G(z i))，同时，根据目标图像生成第一判别结果D(x i)，其中z i为输入至生成网络的低质量图像的数据， G(z i)为生成网络对低质量图像进行优化处理后输出的生成图像的数据，x i为与低质量图像对应的目标图像的数据，D(G(z i))和D(x i)为判别网络输出的生成图像或者目标图像的二分类置信度。进而，可以定义判别器损失函数LossD和生成器损失函数LossG分别如式(5)-(6)所示：
Figure PCTCN2021096024-appb-000005
Figure PCTCN2021096024-appb-000006
其中，x i为与低质量图像对应的目标图像的数据，z i为输入至生成网络的低质量图像的数据，i为任一低质量图像对，m为低质量图像对的总数量。
在根据第一损失函数优化待训练生成对抗深度神经网络模型的参数时，可以先固定生成网络的参数不变，根据公式(5)对判别网络的参数进行优化，使得判别网络的判别准确率达到预设阈值；然后固定判别网络优化后的参数不变，根据公式(6)对生成网络的参数进行优化，使得生成网络能够生成清晰的优化图像。
在本申请的一个实施例中，生成图像和目标图像无论是在低层次的像素值上还是高层次的抽象特征上，都应当接近，因此为了保证生成图像与目标图像在深层语义上的一致，还可以通过分类网络对生成图像和目标图像进行比对，并根据比对结果构建感知损失函数，进而基于感知损失函数对分类网络和生成网络的参数进行优化。其中，分类网络对生成图像进行处理可以获取第一图像信息，对目标图像进行处理可以获取第二图像信息，根据与各个低质量图像对对应的第一图像信息和第二图像信息能够确定第二损失函数，即感知损失函数。其中根据与各个低质量图像对对应的第一图像信息和第二图像信息确定第二损失函数，具体可以将与各低质量图像对对应的第一图像信息和第二图像信息相减，以获取图像信息差；并根据所有低质量图像对对应的图像信息差构建第二损失函数，第二损失函数的表达式如式(7)所示：
Figure PCTCN2021096024-appb-000007
其中，x i为与低质量图像对应的目标图像的数据，z i为输入至生成网络的低质量图像的数据，G(z i)为生成网络对低质量图像进行处理后输出的生成图像的数据，V(G(z i))为第一图像信息，V(x i)为第二图像信息，i为任一低质量图像对，m为低质量图像对的总数量。
在确定第二损失函数后，可以基于第二损失函数对生成网络的参数进行优化，以使生成网络输出的生成图像与目标图像接近或相同。
在本申请的一个实施例中，为了进一步使得生成图像和目标图像接近或相同，可以对生成图像和目标图像中与同一对象对应的图像信息进行比对，若生成图像和目标图像中对应同一对象的图像信息相近或相同说明生成图像与目标图像相近或相同。在本申请的实施例中，可以通过分割网络对目标图像进行分割，以获取图像中每个对象的位置信息。具体到人脸图像，可以采用分割网络对人脸图像进行分割以获取五官分割区域的位置信息，包 括左右眼睛的位置信息、鼻子的位置信息、嘴巴的位置信息；接着根据五官分割区域的位置信息可以从目标图像、生成图像中确定相应区域的图像信息，具体地可以将生成图像中与各对象的位置信息对应的图像信息作为第一局部图像信息，将目标图像中与各对象的位置信息对应的图像信息作为第二局部图像信息；最后计算第一局部图像信息和第二局部图像信息之间的L1范数，即计算对应同一对象的图像信息的绝对值之和，并根据所有低质量图像对对应的L1范数构建第三损失函数。第三损失函数的表达式如式(8)所示：
Figure PCTCN2021096024-appb-000008
其中，x i为与低质量图像对应的目标图像的数据，z i为输入至生成网络的低质量图像的数据，G(z i)为生成网络对低质量图像进行处理后输出的生成图像的数据，M为分割后各对象区域的位置信息，i为任一低质量图像对，m为低质量图像对的总数量。
在本申请的一个实施例中，多个低质量图像对可以看做一批训练样本，根据该训练样本可以对待训练生成对抗深度神经网络模型进行多轮迭代训练，直至完成预设次数的训练或者各个损失函数趋于收敛。在每轮训练过程中，可以依次通过第一损失函数、第二损失函数和第三损失函数对模型的参数进行优化。
在本申请的一个实施例中，以人脸图像为例，图10示出了根据人脸图像对待训练生成对抗深度神经网络模型进行训练的流程总图，此时目标图像为目标人脸图像，低质量图像为低质量人脸图像。如图10所示，在S1001中，获取目标人脸图像后，对目标人脸图像进行对齐处理，以获取目标人脸对齐图像；在S1002中，对该目标人脸对齐图像进行增加噪声处理和/或模糊化处理，以获取低质量人脸图像；在S1003中，将低质量人脸图像输入至生成网络中，通过生成网络进行处理以输出生成人脸图像，该生成人脸图像即为对低质量人脸图像进行优化后的人脸图像；在S1004中，将生成人脸图像和目标人脸图像配对，并分别输入至后处理网络中的判别网络、分类网络和分割网络，通过各个网络对生成人脸图像和目标人脸图像进行特征提取以确定联合损失函数，该联合损失函数包括与各个网络分别对应的第一损失函数、第二损失函数和第三损失函数，进一步地，根据第一损失函数、第二损失函数和第三损失函数对待训练生成对抗深度神经网络模型的参数进行优化，直至生成人脸图像接近目标人脸图像。训练后的待训练生成对抗深度神经网络模型中的生成网络即可被用于对其它经过对齐处理的待优化人脸图像进行优化，以获得不含噪声且生成有面部细节的清晰化人脸图像，进一步提升了用户体验。
在本申请的一个实施例中，在获取生成网络输出的优化图像之后，可以根据待优化图像中人脸相对于标准正面人脸位置的倾斜角度判断是否对优化图像进行位置复位处理，例如当待优化图像中人脸相对于标准正面人脸位置的倾斜角度较小(从视觉效果来说相差不明显)时，可以不对待优化图像进行任何处理，当待优化图像中人脸相对于标准正面人脸位置的倾斜角度较大(从视觉效果来说相差较明显)时，则对待优化图像进行位置复位处理，例如，可以根据对齐处理时所采用的变换矩阵的逆操作将待优化图像中的各个对象还原到原始的位置、角度，最终获取与原始的待优化图像对应的优化图像。
在本申请的一个实施例中，图11A、图11B、图11C、图11D、图11E、图11F示出了三组采用训练后的生成网络对人脸图像进行优化的界面示意图，如图11A、图11C、图11E所示，展示的为待优化人脸图像，可以看出，待优化人脸图像噪声大、清晰度低，边缘轮廓模糊，在经过生成网络对经对齐的待优化人脸对齐图像进行处理后，可以得到清晰度高、 边缘轮廓清晰、面部细节丰富的人脸图像，分别如图11B、图11D、图11F展示的图像所示。
根据本申请提供的图像优化方法，优化处理后的图像(即优化图像)在保留原始图像的特征的基础上，具有清晰度高、包含图像细节，以及噪声被准确去除的特点，同时，由于在模型训练时采用的目标图像均为清晰度很高的图像，因此根据训练后的生成网络对待优化图像进行优化时，能够加深边缘轮廓信息并且还能一定程度地补全图像中缺失的部分，也就是说，本申请中的图像优化方法还具备一定的图像补全和修复能力。另外，本申请的图像优化方法是基于生成对抗深度神经网络模型进行的，因此在优化图像时耗时较短，并且具备扩展性高，可移植性好的特点。
本领域技术人员可以理解实现上述实施方式的全部或部分步骤被实现为由处理器(包含CPU和GPU)执行的计算机程序。例如，通过GPU实现对上述生成对抗深度神经网络模型的训练，或者基于训练后的生成对抗深度神经网络模型，使用CPU或GPU实现对待优化图像的优化处理等。在该计算机程序被处理器执行时，执行本申请提供的上述方法所限定的上述功能。所述的程序可以存储于一种计算机可读存储介质中，该存储介质可以是只读存储器，磁盘或光盘等。
此外，需要注意的是，上述附图仅是根据本申请示例性实施方式的方法所包括的处理的示意性说明，而不是限制目的。易于理解，上述附图所示的处理并不表明或限制这些处理的时间顺序。另外，也易于理解，这些处理可以是例如在多个模块中同步或异步执行的。
以下介绍本申请的图像优化装置实施例，可以用于执行本申请上述的图像优化方法。
图12示出本申请示例性实施例中图像优化装置的结构示意图。如图12所示，上述图像优化装置1200包括：获取模块1201、对齐模块1202和优化模块1203。
其中：获取模块1201，用于获取待优化图像；对齐模块1202，用于对所述待优化图像进行对齐处理，以获取待优化对齐图像，待优化对齐图像的目标区域中各对象的点分布在标准位置；优化模块1203，用于将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；其中所述生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，所述低质量图像对包括目标图像以及所述目标图像对应的低质量图像。
在本申请的一个实施例，所述标准位置模板为特定区域中各对象的点分布；所述对齐模块1201配置为：检测所述待优化图像中的目标区域，所述目标区域与所述特定区域的类型相同；确定所述目标区域的图像数据与所述标准位置模板之间的变换矩阵；根据所述变换矩阵对所述目标区域的图像进行变换操作，以获取所述待优化对齐图像。
在本申请的一些实施例，所述对齐模块配置为：根据标准位置模板对所述待优化图像进行对齐处理，以获取所述待优化对齐图像。
在本申请的一些实施例，所述标准位置模板为特定区域中各对象的点分布，基于前述方案，所述对齐模块配置为：检测所述待优化图像中的目标区域，所述目标区域与所述特定区域的类型相同；确定所述目标区域的图像数据与所述标准位置模板之间的变换矩阵；根据所述变换矩阵对所述目标区域对应的图像进行变换操作，以获取所述待优化对齐图像。
在本申请的一个实施例，所述图像优化装置1200还包括：低质量图像对获取模块，用于获取多个所述低质量图像对；生成图像获取模块，用于将每个低质量图像对分别作为目标图像对，将所述目标图像对中的低质量图像输入至待训练生成对抗深度神经网络模型中的生成网络，得到生成图像；损失函数构建模块，用于将所述生成图像和所述目标图像对中的目标图像输入至所述待训练生成对抗深度神经网络模型中的后处理网络，通过所述后处理网络对所述生成图像和所述目标图像对中的目标图像进行处理以构建所述联合损失函数；模型参数调整模块，用于根据所述联合损失函数优化所述待训练生成对抗深度神经网络模型的参数，得到所述生成网络。
在本申请的一个实施例，所述低质量图像对获取模块配置为：获取多个目标图像，对所述多个目标图像分别进行对齐处理，以获取多个对齐图像；对所述多个对齐图像分别进行低质量化处理，以获取与每个所述目标图像分别对应的低质量图像；根据所述目标图像和所述目标图像对应的低质量图像形成所述低质量图像对。
在本申请的一些实施例，所述低质量化处理包括增加噪声处理和/或模糊化。
在本申请的一个实施例，所述增加噪声处理包括增加高斯噪声、泊松噪声、椒盐噪声中的一种或多种，所述模糊化处理包括均值滤波、高斯滤波、中值滤波、双边滤波、降低分辨率中的一种或多种。
在本申请的一个实施例，所述后处理网络包括判别网络、分类网络和分割网络；所述损失函数构建模块包括：第一损失函数构建单元，用于将所述生成图像和所述目标图像对中的目标图像输入至所述判别网络，获取第一判别结果和第二判别结果，并根据所述第一判别结果和所述第二判别结果构建第一损失函数；第二损失函数构建单元，用于将所述生成图像和所述目标图像对中的目标图像输入至所述分类网络，获取第一图像信息和第二图像信息，并根据所述第一图像信息和所述第二图像信息构建第二损失函数；第三损失函数构建单元，用于将所述生成图像和所述目标图像对中的目标图像输入至所述分割网络，获取第一局部图像信息和第二局部图像信息，并根据所述第一局部图像信息和所述第二局部图像信息构建第三损失函数；联合损失函数构建单元，用于根据所述第一损失函数、所述第二损失函数和所述第三损失函数构建所述联合损失函数。
在本申请的一个实施例，所述第二损失函数构建单元配置为：将每个所述低质量图像对所对应的第一图像信息和第二图像信息相减，以获取图像信息差；根据所有所述低质量图像对所对应的图像信息差构建所述第二损失函数。
在本申请的一个实施例，所述目标图像对中的目标图像和所述生成图像均包括多个对象；基于前述方案，所述第三损失函数构建单元配置为：通过所述分割网络对所述目标图像进行分割，以获取所述目标图像中每个对象的位置信息；将所述生成图像中与每个所述对象的位置信息对应的图像信息作为所述第一局部图像信息，并将所述目标图像中与每个所述对象的位置信息对应的图像信息作为所述第二局部图像信息。
在本申请的一个实施例，所述第三损失函数构建单元配置为：计算所述第一局部图像信息和所述第二局部图像信息之间的L1范数；根据所有所述低质量图像对所对应的L1范数构建所述第三损失函数。
在本申请的一个实施例，所述模型参数调整模块配置为：在每轮训练过程中，依次通过所述第一损失函数、所述第二损失函数和所述第三损失函数优化所述待训练生成对抗深度神经网络模型的参数，得到所述生成网络。
在本申请的一个实施例，所述通过所述第一损失函数优化所述待训练生成对抗深度神经网络模型的参数配置为：固定所述生成网络的参数不变，根据所述第一判别结果和所述第二判别结果优化所述判别网络的参数；固定所述判别网络的优化后的参数不变，根据所述第一判别结果优化所述生成网络的参数。
在本申请的一个实施例，所述生成网络包括：下采样层、残差网络层、上采样层。
上述图像优化装置中各单元的具体细节已经在对应的图像优化方法中进行了详细的描述，因此此处不再赘述。
在本申请的一个实施例中，图像优化装置可以配置于终端设备或者服务器中，当用户通过终端设备请求对其选定的低质量图像进行优化处理时，可以执行上述实施例中的图像优化方法，以得到优化图像。图13A、图13B、图13C、图13D示出了对低质量图像进行优化的界面示意图。如图13A所示，用户可以在终端设备中打开照相功能，界面中“请将摄像头对准脸部”可以指示用户对目标人脸进行拍照；然后对目标人脸进行拍摄以获取低质量人脸图像，在照片浏览界面显示低质量人脸图像，如图13B所示；图13B所示的照片浏览界面包括“重新拍摄”按钮和“优化处理”按钮，当用户通过“重新拍摄”按钮选择重新拍摄，则可以重新拍摄以获取低质量人脸图像，当用户通过“优化处理”按钮在照片浏览界面中针对拍摄得到的低质量人脸图像选择进行优化处理，则可以对低质量人脸图像进行优化处理，选择“优化处理”后，“优化处理”按钮的颜色可以改变，如图13C所示，变成灰色；然后调用图像优化服务对用户拍摄的低质量人脸图像进行优化，并将获得的优化图像返回至优化照片浏览界面，如图13D所示。用户还可以从图册中选择已经拍摄好或下载好的图像进行优化处理，具体处理流程与上述实施例中的图像优化流程相同，在此不再赘述。
图14示出了适于用来实现本申请实施例的电子设备的计算机系统的结构示意图。
需要说明的是，图14示出的电子设备的计算机系统1400仅是一个示例，不应对本申请实施例的功能和使用范围带来任何限制。
如图14所示，计算机系统1400包括处理器1401，其中处理器1401可以包含：图形处理单元(Graphics Processing Unit，GPU)、中央处理单元(Central Processing Unit，CPU)，其可以根据存储在只读存储器(Read-Only Memory，ROM)1402中的程序或者从储存部分1408加载到随机访问存储器(Random Access Memory，RAM)1403中的程序而执行各种适当的动作和处理。在RAM 1403中，还存储有系统操作所需的各种程序和数据。处理器(GPU/CPU)1401、ROM 1402以及RAM 1403通过总线1404彼此相连。输入/输出(Input/Output，I/O)接口1405也连接至总线1404。
计算机系统1400还包括输入部分1406、输出部分1407、通信部分1409、驱动器1410和可拆卸介质1411。
特别地，根据本申请的实施例，下文参考流程图描述的过程可以被实现为计算机软件程序。例如，本申请的实施例包括一种计算机程序产品，其包括承载在计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信部分1409从网络上被下载和安装，和/或从可拆卸介质 1411被安装。在该计算机程序被处理器(GPU/CPU)1401执行时，执行本申请的系统中限定的各种功能。在一些实施例中，计算机系统1400还可以包括AI(Artificial Intelligence，人工智能)处理器，该AI处理器用于处理有关机器学习的计算操作。
需要说明的是，本申请实施例所示的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(Erasable Programmable Read Only Memory，EPROM)、闪存、光纤、便携式紧凑磁盘只读存储器(Compact Disc Read-Only Memory，CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本申请中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本申请中，计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：无线、有线等等，或者上述的任意合适的组合。
描述于本申请实施例中所涉及到的单元可以通过软件的方式实现，也可以通过硬件的方式来实现，所描述的单元也可以设置在处理器中。其中，这些单元的名称在某种情况下并不构成对该单元本身的限定。
作为另一方面，本申请还提供了一种计算机可读介质，该计算机可读介质可以是上述实施例中描述的电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被一个该电子设备执行时，使得该电子设备实现上述实施例中所述的方法。
应当注意，尽管在上文详细描述中提及了用于动作执行的设备的若干模块或者单元，但是这种划分并非强制性的。实际上，根据本申请的实施方式，上文描述的两个或更多模块或者单元的特征和功能可以在一个模块或者单元中具体化。反之，上文描述的一个模块或者单元的特征和功能可以进一步划分为由多个模块或者单元来具体化。
通过以上的实施方式的描述，本领域的技术人员易于理解，这里描述的示例实施方式可以通过软件实现，也可以通过软件结合必要的硬件的方式来实现。因此，根据本申请实施方式的技术方案可以以软件产品的形式体现出来，该软件产品可以存储在一个非易失性存储介质(可以是CD-ROM，U盘，移动硬盘等)中或网络上，包括若干指令以使得一台计算设备(可以是个人计算机、服务器、触控终端、或者网络设备等)执行根据本申请实施方式的方法。

Claims (17)

  1. 一种图像优化方法，所述方法由具有计算处理功能的设备执行，所述方法包括：
    获取待优化图像；
    对所述待优化图像进行对齐处理，以获取待优化对齐图像，所述待优化对齐图像的目标区域中各对象的点分布在标准位置；
    将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；
    其中所述生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，所述低质量图像对包括目标图像以及所述目标图像对应的低质量图像。
  2. 根据权利要求1所述的图像优化方法，对所述待优化图像进行对齐处理，以获取待优化对齐图像，包括：
    根据标准位置模板对所述待优化图像进行对齐处理，以获取所述待优化对齐图像。
  3. 根据权利要求2所述的图像优化方法，所述标准位置模板为特定区域中各对象的点分布，所述根据标准位置模板对所述待优化图像进行对齐处理，以获取所述待优化对齐图像，包括：
    检测所述待优化图像中的目标区域，所述目标区域与所述特定区域的类型相同；
    确定所述目标区域的图像数据与所述标准位置模板之间的变换矩阵；
    根据所述变换矩阵对所述目标区域对应的图像进行变换操作，以获取所述待优化对齐图像。
  4. 根据权利要求1所述的图像优化方法，在将所述待优化对齐图像输入至生成网络中之前，所述方法还包括：
    获取多个所述低质量图像对；
    将每个低质量图像对分别作为目标图像对，将所述目标图像对中的低质量图像输入至所述待训练生成对抗深度神经网络模型中的生成网络，得到生成图像；
    将所述生成图像和所述目标图像对中的目标图像输入至所述待训练生成对抗深度神经网络模型中的后处理网络，通过所述后处理网络对所述生成图像和所述目标图像对中的目标图像进行处理以构建所述联合损失函数；
    根据所述联合损失函数优化所述待训练生成对抗深度神经网络模型的参数，得到所述生成网络。
  5. 根据权利要求4所述的图像优化方法，所述获取多个低质量图像对，包括：
    获取多个目标图像；
    对所述多个目标图像分别进行对齐处理，以获取多个对齐图像；
    对所述多个对齐图像分别进行低质量化处理，以获取与每个所述目标图像分别对应的低质量图像；
    根据所述目标图像和所述目标图像对应的低质量图像形成所述低质量图像对。
  6. 根据权利要求5所述的图像优化方法，所述低质量化处理包括增加噪声处理和/或模糊化。
  7. 根据权利要求6所述的图像优化方法，所述增加噪声处理包括增加高斯噪声、泊松噪声、椒盐噪声中的一种或多种，所述模糊化处理包括均值滤波、高斯滤波、中值滤波、双边滤波、降低分辨率中的一种或多种。
  8. 根据权利要求4所述的图像优化方法，所述后处理网络包括判别网络、分类网络和分割网络，所述通过所述后处理网络对所述生成图像和所述目标图像对中的目标图像进行处理以构建所述联合损失函数，包括：
    将所述生成图像和所述目标图像对中的目标图像输入至所述判别网络，获取第一判别结果和第二判别结果，并根据所述第一判别结果和所述第二判别结果构建第一损失函数；
    将所述生成图像和所述目标图像对中的目标图像输入至所述分类网络，获取第一图像信息和第二图像信息，并根据所述第一图像信息和所述第二图像信息构建第二损失函数；
    将所述生成图像和所述目标图像对中的目标图像输入至所述分割网络，获取第一局部图像信息和第二局部图像信息，并根据所述第一局部图像信息和所述第二局部图像信息构建第三损失函数；
    根据所述第一损失函数、所述第二损失函数和所述第三损失函数构建所述联合损失函数。
  9. 根据权利要求8所述的图像优化方法，所述根据所述第一图像信息和所述第二图像信息构建第二损失函数，包括：
    将每个所述低质量图像对所对应的第一图像信息和第二图像信息相减，以获取图像信息差；
    根据所有所述低质量图像对所对应的图像信息差构建所述第二损失函数。
  10. 根据权利要求8所述的图像优化方法，所述目标图像对中的目标图像和所述生成图像均包括多个对象；
    所述将所述生成图像和所述目标图像对中的目标图像输入至所述分割网络，获取第一局部图像信息和第二局部图像信息，包括：
    通过所述分割网络对所述目标图像进行分割，以获取所述目标图像中每个对象的位置信息；
    将所述生成图像中与每个所述对象的位置信息对应的图像信息作为所述第一局部图像信息，并将所述目标图像中与每个所述对象的位置信息对应的图像信息作为所述第二局部图像信息。
  11. 根据权利要求10所述的图像优化方法，所述根据所述第一局部图像信息和所述第二局部图像信息构建第三损失函数，包括：
    计算所述第一局部图像信息和所述第二局部图像信息之间的L1范数；
    根据所有所述低质量图像对所对应的L1范数构建所述第三损失函数。
  12. 根据权利要求8所述的图像优化方法，所述根据所述联合损失函数优化所述待训练生成对抗深度神经网络模型的参数，得到所述生成网络，包括：
    在每轮训练过程中，依次通过所述第一损失函数、所述第二损失函数和所述第三损失函数优化所述待训练生成对抗深度神经网络模型的参数，得到所述生成网络。
  13. 根据权利要求12所述的图像优化方法，所述通过所述第一损失函数优化所述待训练生成对抗深度神经网络模型的参数，包括：
    固定所述生成网络的参数不变，根据所述第一判别结果和所述第二判别结果优化所述判别网络的参数；
    固定所述判别网络的优化后的参数不变，根据所述第一判别结果优化所述生成网络的参数。
  14. 一种图像优化装置，所述装置部署在具有计算处理功能的设备上，包括：
    获取模块，用于获取待优化图像；
    对齐模块，用于对所述待优化图像进行对齐处理，以获取待优化对齐图像，所述待优化对齐图像的目标区域中各对象的点分布在标准位置；
    优化模块，用于将所述待优化对齐图像输入至生成网络中，通过所述生成网络对所述待优化对齐图像进行特征提取，以获取优化图像；
    其中所述生成网络是根据低质量图像对和联合损失函数对待训练生成对抗深度神经网络模型进行训练得到的，所述低质量图像对包括目标图像以及所述目标图像对应的低质量图像。
  15. 一种计算机可读存储介质，其上存储有计算机程序，所述程序被处理器执行时实现如权利要求1至13中任意一项所述的图像优化方法。
  16. 一种用于图像优化的电子设备，包括：
    一个或多个处理器；
    存储装置，用于存储一个或多个程序，当所述一个或多个程序被所述一个或多个处理器执行时，使得所述一个或多个处理器实现如权利要求1至13中任意一项所述的图像优化方法。
  17. 一种计算机程序产品，当所述计算机程序产品被执行时，用于执行如权利要求1至13中任意一项所述的图像优化方法。
PCT/CN2021/096024 2020-06-28 2021-05-26 图像优化方法、装置、计算机存储介质以及电子设备 Ceased WO2022001509A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
JP2022552468A JP7446457B2 (ja) 2020-06-28 2021-05-26 画像最適化方法及びその装置、コンピュータ記憶媒体、コンピュータプログラム並びに電子機器
EP21832144.6A EP4050511B1 (en) 2020-06-28 2021-05-26 Image optimisation method and apparatus, computer storage medium, and electronic device
US17/735,948 US12175640B2 (en) 2020-06-28 2022-05-03 Image optimization method and apparatus, computer storage medium, and electronic device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010595618.2 2020-06-28
CN202010595618.2A CN111488865B (zh) 2020-06-28 2020-06-28 图像优化方法、装置、计算机存储介质以及电子设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/735,948 Continuation US12175640B2 (en) 2020-06-28 2022-05-03 Image optimization method and apparatus, computer storage medium, and electronic device

Publications (1)

Publication Number Publication Date
WO2022001509A1 true WO2022001509A1 (zh) 2022-01-06

Family

ID=71810596

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/096024 Ceased WO2022001509A1 (zh) 2020-06-28 2021-05-26 图像优化方法、装置、计算机存储介质以及电子设备

Country Status (5)

Country Link
US (1) US12175640B2 (zh)
EP (1) EP4050511B1 (zh)
JP (1) JP7446457B2 (zh)
CN (1) CN111488865B (zh)
WO (1) WO2022001509A1 (zh)

Families Citing this family (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102537207B1 (ko) * 2019-12-30 2023-05-25 포항공과대학교 산학협력단 머신 러닝에 기반한 이미지 처리 방법 및 장치
CN111488865B (zh) 2020-06-28 2020-10-27 腾讯科技（深圳）有限公司 图像优化方法、装置、计算机存储介质以及电子设备
CN112233207B (zh) * 2020-10-16 2025-02-11 北京字跳网络技术有限公司 图像处理方法、装置、设备和计算机可读介质
CN112488944A (zh) * 2020-12-02 2021-03-12 北京字跳网络技术有限公司 样本生成和模型训练方法、装置、设备和计算机可读介质
CN112651948B (zh) * 2020-12-30 2022-04-12 重庆科技学院 一种基于机器视觉的青蒿素萃取智能跟踪与识别方法
CN115131218A (zh) * 2021-03-25 2022-09-30 腾讯科技（深圳）有限公司 图像处理方法、装置、计算机可读介质及电子设备
CN113177982B (zh) * 2021-04-16 2023-03-10 杭州睿影科技有限公司 安检图像数据的处理方法、装置、设备及系统
CN113344832A (zh) * 2021-05-28 2021-09-03 杭州睿胜软件有限公司 图像处理方法及装置、电子设备和存储介质
CN113298807A (zh) * 2021-06-22 2021-08-24 北京航空航天大学 一种计算机断层扫描图像处理方法及装置
CN114299555A (zh) * 2022-01-27 2022-04-08 敦泰电子（深圳）有限公司 指纹识别方法、指纹模组及电子设备
US12020364B1 (en) * 2022-04-07 2024-06-25 Bentley Systems, Incorporated Systems, methods, and media for modifying the coloring of images utilizing machine learning
CN114937187B (zh) * 2022-06-16 2026-03-20 京东科技信息技术有限公司 一种图像优化方法、装置、设备和存储介质
CN115147314B (zh) * 2022-09-02 2022-11-29 腾讯科技（深圳）有限公司 图像处理方法、装置、设备以及存储介质
CN117036180A (zh) * 2022-10-13 2023-11-10 腾讯科技（深圳）有限公司 图像优化方法、装置、电子设备、介质和程序产品
CN115619672B (zh) * 2022-10-20 2026-02-06 深圳前海微众银行股份有限公司 一种图像处理方法、装置、设备及存储介质
CN115689923A (zh) * 2022-10-27 2023-02-03 佛山读图科技有限公司 低剂量ct图像降噪系统与降噪方法
CN116363092A (zh) * 2023-03-27 2023-06-30 中国船舶集团有限公司第七〇九研究所 一种基于神经网络的人群计数方法及装置
CN116363013A (zh) * 2023-04-06 2023-06-30 深圳市威富视界有限公司 图像处理方法及装置
CN116385308B (zh) * 2023-04-15 2024-05-07 广州海至亚传媒科技有限公司 联合图像处理优化策略选择系统
CN116704286B (zh) * 2023-06-02 2024-11-22 中国科学技术大学 一种高密度采样振动图像处理方法及存储介质
CN116977214B (zh) * 2023-07-21 2024-08-06 萱闱(北京)生物科技有限公司 图像优化方法、装置、介质和计算设备
CN118095101A (zh) * 2024-04-17 2024-05-28 中国石油大学(华东) 一种基于机器学习的太赫兹超表面散射参数快速正演方法
CN118097122B (zh) * 2024-04-24 2024-11-22 中南大学 地下目标图像的识别方法、装置、电子设备及存储介质
CN118394869B (zh) * 2024-05-13 2025-01-24 临沂维测工程测绘有限公司 一种国土空间规划用地理信息采集方法及系统
CN118587206B (zh) * 2024-07-30 2024-11-22 维克多精密工业(深圳)有限公司 基于人工智能的模具结构健康监测方法及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108537743A (zh) * 2018-03-13 2018-09-14 杭州电子科技大学 一种基于生成对抗网络的面部图像增强方法
CN109376582A (zh) * 2018-09-04 2019-02-22 电子科技大学 一种基于生成对抗网络的交互式人脸卡通方法
US20190114748A1 (en) * 2017-10-16 2019-04-18 Adobe Systems Incorporated Digital Image Completion Using Deep Learning
CN111488865A (zh) * 2020-06-28 2020-08-04 腾讯科技（深圳）有限公司 图像优化方法、装置、计算机存储介质以及电子设备

Family Cites Families (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5359465B2 (ja) 2009-03-31 2013-12-04 ソニー株式会社 固体撮像装置、固体撮像装置の信号処理方法および撮像装置
CN104318603A (zh) * 2014-09-12 2015-01-28 上海明穆电子科技有限公司 从手机相册调取照片生成3d模型的方法及系统
CN107103590B (zh) * 2017-03-22 2019-10-18 华南理工大学 一种基于深度卷积对抗生成网络的图像反射去除方法
US10565758B2 (en) * 2017-06-14 2020-02-18 Adobe Inc. Neural face editing with intrinsic image disentangling
CN107481188A (zh) * 2017-06-23 2017-12-15 珠海经济特区远宏科技有限公司 一种图像超分辨率重构方法
US11011275B2 (en) 2018-02-12 2021-05-18 Ai.Skopy, Inc. System and method for diagnosing gastrointestinal neoplasm
US10825219B2 (en) * 2018-03-22 2020-11-03 Northeastern University Segmentation guided image generation with adversarial networks
CN108520503B (zh) * 2018-04-13 2020-12-22 湘潭大学 一种基于自编码器和生成对抗网络修复人脸缺损图像的方法
US10284432B1 (en) 2018-07-03 2019-05-07 Kabushiki Kaisha Ubitus Method for enhancing quality of media transmitted via network
CN109685724B (zh) * 2018-11-13 2020-04-03 天津大学 一种基于深度学习的对称感知人脸图像补全方法
CN109615582B (zh) * 2018-11-30 2023-09-01 北京工业大学 一种基于属性描述生成对抗网络的人脸图像超分辨率重建方法
CN109685072B (zh) * 2018-12-22 2021-05-14 北京工业大学 一种基于生成对抗网络的复合降质图像高质量重建方法
CN110349102B (zh) * 2019-06-27 2025-10-10 腾讯科技（深圳）有限公司 图像美化的处理方法、图像美化的处理装置以及电子设备
CN110363116B (zh) * 2019-06-28 2021-07-23 上海交通大学 基于gld-gan的不规则人脸矫正方法、系统及介质
CN110472566B (zh) * 2019-08-14 2022-04-26 旭辉卓越健康信息科技有限公司 一种高精度的模糊人脸识别方法
CN111080527B (zh) * 2019-12-20 2023-12-05 北京金山云网络技术有限公司 一种图像超分辨率的方法、装置、电子设备及存储介质
CN111126307B (zh) * 2019-12-26 2023-12-12 东南大学 联合稀疏表示神经网络的小样本人脸识别方法
CN111179177B (zh) * 2019-12-31 2024-03-26 深圳市联合视觉创新科技有限公司 图像重建模型训练方法、图像重建方法、设备及介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190114748A1 (en) * 2017-10-16 2019-04-18 Adobe Systems Incorporated Digital Image Completion Using Deep Learning
CN108537743A (zh) * 2018-03-13 2018-09-14 杭州电子科技大学 一种基于生成对抗网络的面部图像增强方法
CN109376582A (zh) * 2018-09-04 2019-02-22 电子科技大学 一种基于生成对抗网络的交互式人脸卡通方法
CN111488865A (zh) * 2020-06-28 2020-08-04 腾讯科技（深圳）有限公司 图像优化方法、装置、计算机存储介质以及电子设备

Also Published As

Publication number Publication date
EP4050511A1 (en) 2022-08-31
CN111488865A (zh) 2020-08-04
EP4050511B1 (en) 2025-10-01
JP7446457B2 (ja) 2024-03-08
US20220261968A1 (en) 2022-08-18
CN111488865B (zh) 2020-10-27
US12175640B2 (en) 2024-12-24
JP2023515654A (ja) 2023-04-13
EP4050511A4 (en) 2022-12-28

Similar Documents

Publication Publication Date Title
WO2022001509A1 (zh) 图像优化方法、装置、计算机存储介质以及电子设备
EP3948764B1 (en) Method and apparatus for training neural network model for enhancing image detail
CN110163080B (zh) 人脸关键点检测方法及装置、存储介质和电子设备
EP3916627A1 (en) Living body detection method based on facial recognition, and electronic device and storage medium
CN111767906B (zh) 人脸检测模型训练方法、人脸检测方法、装置及电子设备
CN106682632B (zh) 用于处理人脸图像的方法和装置
CN110516201A (zh) 图像处理方法、装置、电子设备及存储介质
CN109117755B (zh) 一种人脸活体检测方法、系统和设备
US20240404018A1 (en) Image processing method and apparatus, device, storage medium and program product
CN112581370A (zh) 人脸图像的超分辨率重建模型的训练及重建方法
CN108388889B (zh) 用于分析人脸图像的方法和装置
JP2019219928A (ja) 画像処理装置、画像処理方法、及び画像処理プログラム
CN116977548A (zh) 三维重建方法、装置、设备及计算机可读存储介质
CN113436081B (zh) 数据处理方法、图像增强方法及其模型训练方法
WO2020087434A1 (zh) 一种人脸图像清晰度评价方法及装置
CN115222606A (zh) 图像处理方法、装置、计算机可读介质及电子设备
CN117036179A (zh) 图像处理方法、装置、存储介质及计算机设备
CN115375565B (zh) 一种图像中特定形状噪声去除方法、装置及计算机设备
CN114387315B (zh) 图像处理模型训练、图像处理方法、装置、设备及介质
CN117635838A (zh) 三维人脸重建方法、设备、存储介质及装置
CN119646786B (zh) 在线教育平台用户认证方法及系统
Yang et al. An end‐to‐end perceptual enhancement method for UHD portrait images
HK40027974A (zh) 图像优化方法、装置、计算机存储介质以及电子设备
HK40027974B (zh) 图像优化方法、装置、计算机存储介质以及电子设备
CN117152816B (zh) 一种基于卷积神经网络的密集小型人脸检测方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21832144

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2021832144

Country of ref document: EP

Effective date: 20220523

ENP Entry into the national phase

Ref document number: 2022552468

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

WWG Wipo information: grant in national office

Ref document number: 2021832144

Country of ref document: EP

WWG Wipo information: grant in national office

Ref document number: 202237036773

Country of ref document: IN