WO2020052352A1 - 用于对车辆损伤图像进行损伤分割的方法及装置 - Google Patents

用于对车辆损伤图像进行损伤分割的方法及装置 Download PDF

Info

Publication number
WO2020052352A1
WO2020052352A1 PCT/CN2019/097904 CN2019097904W WO2020052352A1 WO 2020052352 A1 WO2020052352 A1 WO 2020052352A1 CN 2019097904 W CN2019097904 W CN 2019097904W WO 2020052352 A1 WO2020052352 A1 WO 2020052352A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
segmentation
category
samples
detection
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/CN2019/097904
Other languages
English (en)
French (fr)
Inventor
徐娟
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.)
Alibaba Group Holding Ltd
Original Assignee
Alibaba Group Holding 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 Alibaba Group Holding Ltd filed Critical Alibaba Group Holding Ltd
Priority to EP19858832.9A priority Critical patent/EP3852061B1/en
Priority to SG11202100117XA priority patent/SG11202100117XA/en
Publication of WO2020052352A1 publication Critical patent/WO2020052352A1/zh
Priority to US17/154,361 priority patent/US11042978B2/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G06T7/001—Industrial image inspection using an image reference approach
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • 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
    • G06N3/08—Learning methods
    • 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
    • G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-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
    • G06N3/09—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
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/10—Office automation; Time management
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G06Q40/08—Insurance
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/174—Segmentation; Edge detection involving the use of two or more images
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • 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
    • G06T2207/10—Image acquisition modality
    • G06T2207/10004—Still image; Photographic image
    • 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
    • G06T2207/20—Special algorithmic details
    • G06T2207/20112—Image segmentation details
    • G06T2207/20132—Image cropping
    • 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
    • G06T2207/20112—Image segmentation details
    • G06T2207/20164—Salient point detection; Corner detection
    • 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
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30156—Vehicle coating
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07C—TIME OR ATTENDANCE REGISTERS; REGISTERING OR INDICATING THE WORKING OF MACHINES; GENERATING RANDOM NUMBERS; VOTING OR LOTTERY APPARATUS; ARRANGEMENTS, SYSTEMS OR APPARATUS FOR CHECKING NOT PROVIDED FOR ELSEWHERE
    • G07C5/00—Registering or indicating the working of vehicles
    • G07C5/08—Registering or indicating performance data other than driving, working, idle, or waiting time, with or without registering driving, working, idle or waiting time
    • G07C5/0808—Diagnosing performance data

Definitions

  • the embodiments of the present specification relate to the field of vehicle damage determination, and in particular, to a method and device for performing damage segmentation on a vehicle damage image.
  • obtaining pixel-level segmentation results of damaged objects from vehicle damage images is very important to improve the accuracy of identifying damaged objects, and to accurately locate and accurately display damaged objects.
  • the method of manually labeling is used to determine pixel-level damage areas in car damage images.
  • the shape of the exterior damage of the car body includes a large number of discontinuous, irregular scratches, deformations, tears, etc., which makes it difficult to confirm the pixel boundary of the damaged area, so it is difficult to manually mark it.
  • a damage segmentation model is determined, which is used for damage segmentation of a vehicle damage image.
  • a method for damage segmentation of a vehicle damage image includes: obtaining a plurality of segmentation samples labeled with a target detection result and a target segmentation result simultaneously, the target detection result includes a category and a frame of the target object, the target segmentation result includes a contour of the target object, and the plurality of segmentation Multiple categories labeled on the target object in the sample constitute a first category subset; multiple damage detection samples labeled with target detection results but not labeled with target segmentation results are obtained; multiple of the multiple damage detection samples labeled with target objects Categories constitute a second category subset, and the second category subset includes categories that do not belong to the first category subset; using the plurality of segmented samples and the plurality of damage detection samples to train an injury segmentation model, the The damage segmentation model includes target detection parameters for detecting a damaged object and target segmentation parameters for determining the contour of the damaged object; wherein the training damage segmentation model includes:
  • the parameters are input into the weight migration function to obtain target segmentation parameters corresponding to each category, so as to determine the damage segmentation model for damage segmentation of a vehicle damage image.
  • the first category subset includes a first category
  • the determining the weight transfer function includes: obtaining an object detection parameter corresponding to the first category; and detecting an object corresponding to the first category
  • the parameters are input into the initial weight migration function to obtain corresponding prediction target segmentation parameters; based on the prediction target segmentation parameters, a prediction target segmentation result corresponding to the segmentation sample of the first category is determined; at least based on the prediction target segmentation result and the first
  • the target segmentation results labeled in the segmented samples of the categories are adjusted by the initial weight transfer function.
  • the target detection parameters corresponding to the first category include target classification parameters and target frame parameters; and the target detection parameters corresponding to the first category are input into an initial weight.
  • the migration function includes: inputting the target classification parameter and / or the target frame parameter into the initial weight migration function.
  • the weight transfer function is implemented by a convolutional neural network.
  • the plurality of segmented samples includes a plurality of component segmented samples whose target object is a component object.
  • the plurality of segmented samples includes a plurality of damaged segmented samples whose target object is a damaged object.
  • the obtaining a plurality of segmented samples simultaneously labeled with a target detection result and a target segmentation result includes: obtaining a predetermined number of damage detection samples from a damage sample library; based on saliency detection Extracting a plurality of saliency areas from a plurality of frames marked by the predetermined number of damage detection samples, so that a staff member may perform a part of the predetermined number of damage detection samples according to the plurality of saliency areas. Label the segmentation results; determine the number of damage detection samples labeled with the segmentation result as the plurality of damage segmentation samples; and use the damage detection samples other than the partial number of the predetermined number as the plurality of damages Test samples.
  • a method for segmenting a vehicle damage image includes: obtaining a plurality of segmentation samples labeled with a target detection result and a target segmentation result simultaneously, the target detection result includes a category and a frame of the target object, the target segmentation result includes a contour of the target object, and the plurality of segmentation Multiple categories labeled for the target object in the sample constitute a third category subset; obtaining multiple component detection samples labeled with target detection results but not labeled with target segmentation results; multiple component labeled samples with multiple target objects
  • the categories constitute a fourth category subset, and the fourth category subset includes categories that do not belong to the third category subset; using the plurality of segmented samples and the plurality of component detection samples to train a component segmentation model, the The component segmentation model includes target detection parameters for detecting component objects and target segmentation parameters for determining the contours of the component objects; wherein the training component segmentation model includes:
  • the parameters are input into the weight migration function to obtain target segmentation parameters corresponding to each category, thereby determining the component segmentation model for component segmentation of a vehicle damage image.
  • the plurality of segmented samples includes a plurality of component segmented samples whose target object is a component object.
  • a device for segmenting a vehicle damage image includes: a first acquisition unit configured to acquire a plurality of segmentation samples labeled with a target detection result and a target segmentation result at the same time;
  • the target detection result includes the category and frame of the target object, and the target segmentation result includes the outline of the target object;
  • the multiple categories marked on the target object in the plurality of segmentation samples constitute a first category subset;
  • the second acquisition unit configures In order to obtain multiple damage detection samples labeled with target detection results but not labeled with target segmentation results, multiple categories labeled for target objects in the multiple damage detection samples constitute a second category subset, and the second category subset Including a category that does not belong to the first category subset;
  • a training unit configured to use the plurality of segmented samples and the plurality of damage detection samples to train a damage segmentation model that includes a target for detecting a damaged object Detection parameters and target segmentation parameters for determining the contour of the damaged object;
  • a first determination module configured to determine a category set composed of the first category subset and the second category subset based on the target detection results in the plurality of segmented samples and the target detection results in the plurality of damage detection samples.
  • Target detection parameters corresponding to each category in the category
  • a second determination module configured to determine a weight migration function based on the target detection results and the target segmentation results in the plurality of segmented samples, the weight migration function representing a transition from the target detection parameter to the target Mapping of segmentation parameters
  • an input module configured to input target detection parameters corresponding to each category into the weight transfer function to obtain target segmentation parameters corresponding to each category, thereby determining the damage segmentation model for vehicle damage The image is segmented for damage.
  • an apparatus for segmenting a vehicle damage image includes: a first obtaining unit configured to obtain a plurality of segmentation samples labeled with a target detection result and a target segmentation result simultaneously, the target detection result includes a category and a frame of the target object, and the target segmentation result includes a target object Outline; multiple categories labeled on the target object in the multiple segment samples form a third category subset; a second acquisition unit configured to obtain multiple component detection samples labeled with target detection results but not labeled with target segmentation results; The multiple categories marked on the target object in the multiple component detection samples constitute a fourth category subset, and the fourth category subset includes categories that do not belong to the third category subset; a training unit configured to use the multiple Segmentation samples and the plurality of component detection samples, training a component segmentation model, the component segmentation model including target detection parameters for detecting component objects and target segmentation parameters for determining contours of the component objects; wherein the training The units specifically include:
  • a first determining module configured to determine a category set composed of the third category subset and the fourth category subset based on the target detection results in the plurality of segmented samples and the target detection results in the multiple component detection samples.
  • Target detection parameters corresponding to each category in the category a second determination module configured to determine a weight migration function based on the target detection results and the target segmentation results in the plurality of segmented samples, the weight migration function representing a transition from the target detection parameter to the target Mapping of segmentation parameters; an input module configured to input target detection parameters corresponding to each category into the weight migration function to obtain target segmentation parameters corresponding to each category, thereby determining the component segmentation model for vehicle damage Image segmentation.
  • a computer-readable storage medium having stored thereon a computer program, which when executed in a computer, causes the computer to execute the method described in the first aspect or the second aspect.
  • a computing device including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the first aspect or the first aspect is implemented. The method described in the second aspect.
  • a segmentation sample corresponding to a first category subset and labeled with a target detection result and a target segmentation result at the same time, and a subcategory with a second category Set the corresponding damage detection samples labeled with target detection results but not labeled with target segmentation results to determine the target detection parameters and target segmentation parameters corresponding to each category in the category set consisting of the first category subset and the second category subset, so that Determine the damage segmentation model for damage segmentation of vehicle damage images.
  • Figure 1 shows a partial photograph of a vehicle according to an example
  • FIG. 2 is a schematic diagram of an interface change of a fixed loss client according to an example
  • FIG. 3 shows a flowchart of a method for damage segmentation of a vehicle damage image according to an embodiment
  • FIG. 4 is a schematic flowchart of a method for obtaining a damaged segmented sample according to an embodiment
  • FIG. 5 illustrates a schematic diagram of extracting a damaged object based on saliency detection according to an embodiment
  • FIG. 6 shows a flowchart of a training method of a damage segmentation model according to an embodiment
  • FIG. 7 illustrates a Mask-R-CNN architecture diagram based on transfer learning according to an embodiment
  • FIG. 8 illustrates a flowchart of a method for segmenting a vehicle damage image according to an embodiment
  • FIG. 9 shows a structural diagram of an apparatus for damage segmentation of a vehicle damage image according to an embodiment
  • FIG. 10 shows a structural diagram of a device for segmenting a vehicle damage image according to an embodiment.
  • An embodiment of the present specification discloses a method for performing damage segmentation on a vehicle damage image, wherein the damage segmentation refers to region extraction of a damaged object having a target category and a clear boundary in the vehicle injury image, and the region extraction can be expressed as determining the damage The outline of the object.
  • the method specifically includes determining a damage segmentation model, so that a vehicle damage image can be input into the model, thereby obtaining a target category (hereinafter referred to as a damage category) and a segmentation contour of the damage object.
  • a damage category a target category
  • the damage segmentation model can be applied to the fixed loss client provided to the user.
  • a user can take a scene photo, such as a partial picture of a vehicle shown in FIG.
  • the damage determination client can use the damage segmentation model to determine the vehicle damage information corresponding to the scene photos.
  • the vehicle damage category can be determined to be moderate scratches and the outline of scratch damage.
  • a maintenance plan corresponding to the damage information and a related compensation amount can also be given, for example, the maintenance plan is: paint repair, and the warranty compensation amount: 120 yuan.
  • an embodiment of the present specification discloses a method for performing damage segmentation on a vehicle damage image.
  • the transfer learning method is used to determine the damage segmentation model.
  • the transfer learning can be understood as using existing Knowledge to learn new knowledge, the core is to find the similarity between existing knowledge and new knowledge.
  • a large number of segmented samples corresponding to non-damaged objects with similar characteristics to the damaged objects may be obtained, and these segmented samples are usually relatively easy to obtain, and a damage segmentation model is determined.
  • a segmented sample with a target object as the component object can be used. Because the contour features of the component are similar to the contour features of the damage, and because the contour of the part is more regular than the damaged contour, a large number of segmented samples labeled with the segmentation result are compared Easy.
  • multiple segmented samples with similar characteristics to vehicle damage, and multiple damage detection samples can be obtained, where the segmented samples are labeled with target detection results (including the type and border of the target object) and targets.
  • Segmentation results, and multiple categories corresponding to multiple segmented samples constitute category set A
  • damage detection samples are labeled with target detection results but not target segmentation results
  • multiple categories corresponding to multiple damage detection samples constitute category set B
  • the category set B has categories that do not belong to the category set A; then, determine the relationship between the target detection result and the target segmentation result corresponding to each category in the segmented sample, and then migrate this relationship to the damage detection sample, and then A target segmentation parameter corresponding to each category in the category set B can be determined based on this relationship, and is used to segment the damage objects corresponding to the category set B in the vehicle damage image.
  • FIG. 3 shows a flowchart of a method for damage segmentation of a vehicle damage image according to an embodiment.
  • the method may be performed by a device having a processing capability: a server or a system or a device.
  • the method flow includes the following steps: Step S310, obtaining a plurality of segmentation samples labeled with a target detection result and a target segmentation result at the same time, and a plurality of categories labeled on the target object in the segmentation samples constitute a first category sub
  • Step S320 obtaining multiple damage detection samples labeled with target detection results but not labeled with target segmentation results, and multiple categories labeled with target objects in these detection samples constitute a second category subset, and the second category subset includes A category that does not belong to the first category subset; step S330, using the plurality of segmented samples and the plurality of damage detection samples to train a damage segmentation model, the model including target detection parameters for detecting a damaged object and Determine the target segmentation parameters of the damaged object contour.
  • step S310 a plurality of segmentation samples are simultaneously labeled with a target detection result and a target segmentation result.
  • the target detection result includes the category and frame of the target object
  • the target segmentation result includes the outline of the target object, or a mask corresponding to the outline of the target object.
  • the target object is an object with similar or similar characteristics to the vehicle injury, so that the trained injury segmentation model can obtain better segmentation results.
  • the target object may include a vehicle component
  • the multiple segmented samples may include multiple component segmented samples.
  • the multiple component segmentation samples may include component segmentation samples obtained based on manual labeling. It should be noted that, because the contours of the vehicle components in the vehicle damage image are relatively regular, it is highly feasible to manually label the segmentation results of the vehicle components.
  • the plurality of component segmentation samples may further include a component segmentation sample obtained based on the component segmentation model.
  • a component segmentation model can be trained based on the aforementioned manually labeled component segmentation samples, and then a large number of vehicle damage pictures are input into the component segmentation model to obtain more images with the result of component segmentation, and these images are used as Multiple parts split a portion of the sample.
  • the target object may include vehicle damage
  • the multiple segmented samples may include multiple damaged segmented samples. It should be noted that the number of categories corresponding to these damage segmentation samples is less, or in other words, far less than the number of categories corresponding to the damage detection samples that will be mentioned later. Further, as shown in FIG. 4, a damage segmentation sample can be obtained by the following steps:
  • step S41 a predetermined number of damage detection samples are acquired from the damage sample library.
  • the damage detection sample is a sample marked with the target detection result but not labeled with the target segmentation result, that is, the damage detection sample is labeled with the type of the damage object and the frame where the damage object is located, but the damage object is not marked. Segmentation results.
  • damage detection samples it is relatively easy to obtain damage detection samples.
  • the operation of manually marking the damage frame and the damage category in the vehicle damage image is relatively simple. Therefore, a large number of damage detection samples can be obtained based on the manual marking.
  • a corresponding damage detection result may be obtained based on an existing damage detection model and a large number of vehicle damage pictures obtained from an insurance company, and a damage detection sample may be determined according to the damage detection result.
  • the predetermined number may be determined by a staff member based on actual experience.
  • step S42 based on the saliency detection, a plurality of saliency regions are extracted from a plurality of frames marked by the acquired predetermined number of damage detection samples, so that the staff member can perform Partial number of damage detection samples are labeled for segmentation results.
  • saliency detection also called visual attention mechanism
  • saliency detection refers to simulating human visual characteristics through intelligent algorithms and extracting saliency regions (that is, regions of human interest) in the image.
  • saliency regions that is, regions of human interest
  • saliency detection can be achieved.
  • the saliency detection can be performed on the image in the frame to quickly extract the damaged object therein.
  • a saliency metric map shown in FIG. 5 (b) can be obtained, in which the white part is the detected damage object .
  • the saliency area extracted based on the saliency detection may be directly and automatically labeled as the segmentation result of the corresponding damage object.
  • the frame may include significant objects other than damage objects such as dust and stains, and on the other hand, due to the variety of vehicle damage, and including a large number of discontinuous damage, or a relatively minor degree Damage, so for some damage detection samples, the saliency detection may not be able to completely and accurately extract the area covered by the damage object, so in another specific embodiment, multiple extracted saliency
  • the area is provided to the staff, so that the staff can mark the segmentation results of a predetermined number of damage detection samples according to multiple salient areas. For example, it can include filtering or correcting the salient areas to obtain the segmentation result. More accurate damage segmentation samples.
  • step S43 the obtained partial number of damage detection samples is used as a plurality of damage segmentation samples.
  • the obtained multiple segmented samples may include the multiple component segmented samples described above, and / or multiple lesion segmented samples obtained based on the significance detection.
  • the plurality of segmented samples may also include a small number of damaged segmented samples obtained based on manual labeling.
  • the plurality of segmented samples may further include other segmented samples with similar characteristics to the damaged object, for example, the segmented sample in which the target object is a plant leaf.
  • step S320 a plurality of damage detection samples labeled with target detection results but not labeled with target segmentation results are obtained.
  • the multiple categories marked on the target object in the multiple damage detection samples constitute a second category subset, and the second category subset includes categories that do not belong to the first category subset.
  • the multiple segmented samples obtained in step S310 include component segmented samples whose target object is a component object, and the target object of the damage detection sample is a damaged object.
  • the category in the first category subset is the component category.
  • the category in the second category subset is the damage category, which is different from the component category.
  • the multiple segmented samples obtained in step S310 include a partial number of damaged segmented samples. As can be seen from the foregoing, the damage categories corresponding to these damaged segmented samples are relatively limited, so the damage categories corresponding to the damaged segmented samples Is less, or much less than, the number of injury categories included in the second category subset.
  • the remaining (eg, a predetermined number minus a partial number) of damage detection samples may be used as Part of a plurality of damage detection samples obtained in this step.
  • multiple segmented samples can be acquired in step S310, and multiple damage detection samples can be acquired in step S320.
  • a plurality of segmented samples and a plurality of damage detection samples are used to train a damage segmentation model.
  • the damage segmentation model includes target detection parameters for detecting a damaged object and target segmentation parameters for determining a contour of the damaged object. .
  • training the injury segmentation model may include the following steps:
  • step S61 based on the target detection results in a plurality of segmented samples and the target detection results in a plurality of damage detection samples, the corresponding categories in the category set composed of the first category subset and the second category subset are determined.
  • Target detection parameters based on the target detection results in a plurality of segmented samples and the target detection results in a plurality of damage detection samples.
  • an existing target detection algorithm such as Faster R-CNN, R-FCN, or SSD, may be used to determine the target detection parameters corresponding to each category in the category set.
  • a weight migration function is determined based on the target detection results and the target segmentation results in the plurality of segmented samples, and the weight migration function represents a mapping from the target detection parameter to the target segmentation parameter.
  • the target detection parameter and the target segmentation parameter are determined separately, and then the mapping relationship between the two is determined as a weight transfer function by a mathematical method. Specifically, first, the target detection parameters corresponding to each category in the first category subset can be determined according to the target detection results in the segmented samples, and the target segmentation corresponding to each category in the first category category subset can be determined based on the target segmentation results. Parameters; then, a weight transfer function can be determined based on the target detection parameters and target segmentation parameters corresponding to each category.
  • the weight transfer function is determined by training.
  • an initial weight transfer function is determined (ie, preliminary function parameters are determined), and then target detection parameters are used as inputs, target segmentation results are used as labels, and function weight adjustments are used to obtain weight transfer functions.
  • a certain category hereinafter referred to as the first category included in the first category subset is used as an example to describe some steps of the training process.
  • the initial weight transfer function is adjusted. In this way, the weight transfer function is trained.
  • the target detection parameters corresponding to the first category include target classification parameters and target frame parameters. Accordingly, inputting the target detection parameters corresponding to the first category into the initial weight migration function may include: The classification parameters, and / or, the target frame parameters are entered into the initial weight transfer function.
  • the target detection parameter may be a combination of a target classification parameter and a target frame parameter.
  • the weight transfer function can be expressed as
  • ⁇ det represents a target detection function of any category
  • ⁇ seg represents a target segmentation parameter corresponding to the same category as ⁇ det
  • ⁇ is a learning parameter independent of the category.
  • ⁇ can be set to ⁇ 0 , and the specific value can be randomly selected.
  • the weight transfer function may be implemented by a convolutional neural network, that is, the operation of the weight transfer function is implemented by a combination of neurons of the neural network. Accordingly, the training of the weight transfer function includes adjusting and determining the operation parameters of the neurons in the neural network and the connection weight parameters between the neurons.
  • adjusting the initial weight migration function may include determining that the target segmentation result corresponds to the target segmentation result based on the prediction target segmentation result corresponding to one or more categories in the first category set and the target segmentation result labeled in the segmentation sample. Loss function, adjust the initial weight transfer function by means of error back propagation or gradient descent.
  • the target detection parameters corresponding to the first category c are first obtained Then, change Enter the initial weight migration function represented by:
  • determining the weight transfer function includes determining a parameter value of ⁇ in formula (1).
  • step S61 and step S62 can be performed independently at the same time, or can be performed independently one after the other, and step S62 can be performed simultaneously as a branch of step S61 to save resource overhead.
  • the weight migration function can be determined, and this function represents the mapping from the target detection parameter to the target segmentation parameter. Then, in step S63, the target detection parameters corresponding to each category in the category set determined in step S61 are input into a weight migration function to obtain target segmentation parameters corresponding to each category, thereby determining a damage segmentation model for performing a vehicle damage image. Damage segmentation.
  • target segmentation parameters corresponding to each category in the category set composed of the first category subset and the second category subset can be obtained, and can be used to segment target objects included in the image corresponding to each category in the category set.
  • This also means that for some categories in the second category subset, in the case where only corresponding detection samples exist in the training sample set, and no corresponding divided samples exist, the method provided in the embodiment of the present specification is adopted, Corresponding target segmentation parameters can also be obtained, which are used to segment the vehicle damage images corresponding to these categories.
  • step S330 further describes step S330 with a specific example.
  • a Mask-R-CNN architecture diagram based on transfer learning is shown, where the first category subset is set A, the second category subset is set B, and ABB represents set A and set The union of B, that is, the aforementioned set of categories.
  • Box features are extracted from the aggregated area features through the box head. Based on the box features and box detection parameters (box weights), they can be expressed as w det to determine the prediction results (box predictions). ), And based on the prediction detection result and the label detection result (box lables in A ⁇ B) corresponding to this image, calculate a loss function (box loss) corresponding to the detection result to adjust the w det corresponding to this image category;
  • the target segmentation parameters (mask weights) corresponding to the category are determined based on the weight transfer function, which can be expressed as w seg , where the weight transfer function is used to represent the mapping between the target detection parameters and the target segmentation parameters. relationship.
  • a weight transfer function can be determined.
  • this image is an image corresponding to the detection sample
  • the processing is performed the same as in 1) above, and further, target detection parameters corresponding to the category of this image can be obtained.
  • the weight transfer function can be determined based on multiple segmented samples corresponding to category set A, and at the same time, based on multiple segmented samples corresponding to category set A and multiple damage detection samples corresponding to category set B, A ⁇ can be determined.
  • Target detection parameters corresponding to each category in B can be determined.
  • the method for damage segmentation can be based on segmented samples corresponding to the first category subset and labeled with the target detection result and the target segmentation result, and corresponding to the second category subset.
  • Damage detection samples labeled with target detection results but not labeled with target segmentation results determine target detection parameters and target segmentation parameters corresponding to each category in the category set consisting of the first category subset and the second category subset to determine the damage Segmentation model for damage segmentation of vehicle damage images.
  • FIG. 8 shows a flowchart of a method for damage segmentation of a vehicle damage image according to an embodiment.
  • the method may be performed by a device having a processing capability: a server or a system or a device. As shown in FIG.
  • Step S810 Obtain multiple segmented samples labeled with a target detection result and a target segmentation result at the same time, and among the segmented samples, multiple categories labeled with the target object constitute a third category sub-group
  • Step S820 obtaining a plurality of component detection samples labeled with target detection results but not labeled with target segmentation results, and multiple categories labeled with target objects in these detection samples constitute a fourth category subset, and the fourth category subset includes A category that does not belong to the third category subset
  • step S830 using multiple segmented samples and multiple component detection samples to train a component segmentation model, the model including target detection parameters for detecting component objects and for determining the components Object segmentation parameters for object contours.
  • step S810 a plurality of segmentation samples labeled with a target detection result and a target segmentation result are acquired simultaneously, and a plurality of categories labeled with a target object in the segmentation samples constitute a third category subset.
  • the plurality of segmented samples may include segmented samples having similar contour features as the component, such as leaf segmented samples of various plants.
  • the multiple segmentation samples include multiple component segmentation samples whose target object is a component object.
  • the categories corresponding to the obtained multiple component segmentation samples are usually only a part of the vehicle components. In one example, they may account for 1/3 or 1 of all vehicle component categories. / 4.
  • the corresponding component detection samples are easier to obtain, and the component segmented samples marked with the segmentation results are more difficult to obtain.
  • the possibility of damage in a vehicle accident is relatively small, so there are limited pictures of vehicle damage associated with it.
  • the cost of labeling the segmentation results is high, so the segmentation results of such components The possibility of manual labeling is less, which makes it difficult to obtain segmented samples of these components.
  • steps S810 to S830 reference may also be made to the foregoing description of steps S310 to S330, and details are not described herein.
  • the method for component segmentation provided by the embodiment of the present specification can be based on a segmentation sample corresponding to the third category subset and labeled with the target detection result and the target segmentation result, and corresponding to the fourth category subset.
  • Component detection samples labeled with target detection results but not labeled with segmentation results determine target detection parameters and target segmentation parameters corresponding to each category in the category set consisting of the first category subset and the second category subset, thereby determining the component Segmentation model for segmenting parts of a vehicle damage image.
  • FIG. 9 shows a structural diagram of a device for damage segmentation of a vehicle damage image according to an embodiment. As shown in Figure 9, the device includes:
  • the first obtaining unit 910 is configured to obtain a plurality of segmentation samples labeled with a target detection result and a target segmentation result simultaneously, where the target detection result includes a category and a frame of the target object, and the target segmentation result includes a contour of the target object;
  • the multiple categories labeled on the target object in the multiple segmented samples constitute a first category subset;
  • a second obtaining unit 920 configured to obtain a plurality of damage detection samples labeled with a target detection result but not labeled with a target segmentation result; a plurality of categories labeled on the target object in the plurality of damage detection samples constitute a second category subset, The second category subset includes categories that do not belong to the first category subset;
  • a training unit 930 is configured to use the plurality of segmented samples and the plurality of damage detection samples to train a damage segmentation model, where the damage segmentation model includes a target detection parameter for detecting a damaged object and for determining the damaged object Contour segmentation parameters; the trained unit 930 specifically includes:
  • the first determining module 931 is configured to determine a category formed by the first category subset and the second category subset based on the target detection results in the plurality of segmented samples and the target detection results in the plurality of damage detection samples. Target detection parameters corresponding to each category in the set;
  • a second determining module 932 configured to determine a weight migration function based on the target detection result and the target segmentation result in the plurality of segmented samples, where the weight migration function represents a mapping from a target detection parameter to a target segmentation parameter;
  • An input module 933 is configured to input target detection parameters corresponding to the respective categories into the weight migration function to obtain target segmentation parameters corresponding to each category, thereby determining the damage segmentation model for performing damage segmentation on a vehicle damage image.
  • the first category subset includes a first category
  • the second determining module 932 is specifically configured to:
  • the target detection parameters corresponding to the first category include a target classification parameter and a target frame parameter;
  • the input module 933 is specifically configured as:
  • the weight transfer function is implemented by a convolutional neural network.
  • the plurality of segmented samples include a plurality of component segmented samples whose target object is a component object.
  • the plurality of segmented samples include a plurality of damaged segmented samples whose target object is a damaged object.
  • the first obtaining unit 910 is specifically configured as:
  • a plurality of saliency regions are extracted from a plurality of frames marked by the predetermined number of damage detection samples, so that a staff member may perform a partial number of the predetermined number according to the plurality of saliency regions. Segmentation results of the damage detection samples;
  • a part number of damage detection samples marked with a segmentation result is determined as the plurality of damage segmentation samples; and a damage detection sample other than a part number of the predetermined number is used as the plurality of damage detection samples.
  • the device for damage segmentation provided by the embodiments of the present specification can be based on segmented samples corresponding to the first category subset and labeled with the target detection result and the target segmentation result, and corresponding to the second category subset.
  • Damage detection samples labeled with target detection results but not labeled with target segmentation results determine target detection parameters and target segmentation parameters corresponding to each category in the category set consisting of the first category subset and the second category subset to determine the damage Segmentation model for damage segmentation of vehicle damage images.
  • FIG. 10 shows a structural diagram of a device for segmenting a vehicle damage image according to an embodiment. As shown in FIG. 10, the device includes:
  • the first obtaining unit 1010 is configured to obtain a plurality of segmentation samples labeled with a target detection result and a target segmentation result simultaneously, where the target detection result includes a category and a frame of the target object, and the target segmentation result includes a contour of the target object;
  • the multiple categories labeled for the target object in the multiple segmented samples constitute a third category subset;
  • the second obtaining unit 1020 is configured to obtain multiple component detection samples labeled with target detection results but not labeled with target segmentation results; multiple categories labeled with target objects in the multiple component detection samples constitute a fourth category subset, The fourth category subset includes categories that do not belong to the third category subset;
  • a training unit 1030 configured to use the plurality of segmented samples and the plurality of component detection samples to train a component segmentation model, the component segmentation model including a target detection parameter for detecting a component object and for determining the component object Contour segmentation parameters; the training unit 1030 specifically includes:
  • the first determining module 1031 is configured to determine a category composed of the third category subset and the fourth category subset based on the target detection results in the plurality of segmented samples and the target detection results in the plurality of component detection samples.
  • a second determining module 1032 configured to determine a weight migration function based on the target detection result and the target segmentation result in the plurality of segmented samples, where the weight migration function represents a mapping from a target detection parameter to a target segmentation parameter;
  • An input module 1033 is configured to input target detection parameters corresponding to the respective categories into the weight transfer function to obtain target segmentation parameters corresponding to each category, thereby determining the component segmentation model for component segmentation of a vehicle damage image. .
  • the plurality of segmented samples includes a plurality of component segmented samples whose target object is a component object.
  • the device for component segmentation provided by the embodiment of the present specification can be based on segmentation samples corresponding to the third category subset and labeled with the target detection result and the target segmentation result, and corresponding to the fourth category subset Component detection samples labeled with target detection results but not labeled with segmentation results, determine target detection parameters and target segmentation parameters corresponding to each category in the category set consisting of the first category subset and the second category subset, thereby determining the component Segmentation model for segmenting parts of a vehicle damage image.
  • a computer-readable storage medium in which a computer program is stored, and when the computer program is executed in the computer, the computer is executed in combination with FIG. 3, FIG. 4, and FIG. Or the method described in Figure 8.
  • a computing device which includes a memory and a processor, where the executable code is stored in the memory, and when the processor executes the executable code, the combination of FIG. 3 and FIG. 4 is implemented. , Figure 6 or Figure 8.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Business, Economics & Management (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Accounting & Taxation (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Quality & Reliability (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Economics (AREA)
  • Technology Law (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Human Resources & Organizations (AREA)
  • Development Economics (AREA)
  • Operations Research (AREA)
  • Tourism & Hospitality (AREA)
  • Image Analysis (AREA)

Abstract

本说明书实施例提供基于迁移学习的用于部件分割和损伤分割的方法，具体包括：第一方面，可以基于具有丰富的检测和分割数据的高频部件类别学习一种迁移方法，然后对于只有检测数据没有丰富分割数据的低频部件类别可以通过其检测框作用于此迁移方法而获得对应的分割结果；第二方面，可以基于具有丰富的检测和分割数据的高频损伤类别学习一种检测到分割迁移方法，然后对于只有检测数据没有丰富分割数据的低频损伤类别可以通过其检测框作用于此迁移方法而获得对应的分割结果；第三方面，可以通过部件到损伤的跨域迁移学习获得损伤的分割结果。通过这些方法，可以解决长尾的部件类别和损伤类别的分割数据获取困难的问题。

Description

用于对车辆损伤图像进行损伤分割的方法及装置 技术领域
本说明书实施例涉及车辆定损领域，具体地，涉及一种用于对车辆损伤图像进行损伤分割的方法及装置。
背景技术
在车险理赔场景中，从车辆损伤图像中获取损伤对象的像素级分割结果，对提升识别损伤对象的准确度，以及对损伤对象的精准定位和精准展示来说，都十分重要。
目前，主要是采用人工标注的方法确定出车损图像中的像素级损伤区域。然而，车体外观损伤的形状包括大量不连续、不规则的刮擦，变形，撕裂等，导致损伤区域的像素点边界难以确认，因而进行人工标注的难度较大。
因此，需要一种更加有效地方法，可以快速、准确地实现对车辆损伤图像的损伤分割。
发明内容
在本说明书描述的一种用于对车辆损伤图像进行损伤分割的方法中，基于迁移学习的思想，确定出损伤分割模型，用于对车辆损伤图像进行损伤分割。
根据第一方面，提供一种用于对车辆损伤图像进行损伤分割的方法。该方法包括：获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第一类别子集；获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本；所述多个损伤检测样本中对目标对象标注的多个类别构成第二类别子集，所述第二类别子集中包括不属于第一类别子集的类别；利用所述多个分割样本和所述多个损伤检测样本，训练损伤分割模型，所述损伤分割模型包括用于检测损伤对象的目标检测参数和用于确定所述损伤对象轮廓的目标分割参数；其中所述训练损伤分割模型包括：
基于所述多个分割样本中的目标检测结果和多个损伤检测样本中的目标检测结果，确定与所述第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测 参数；基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述损伤分割模型，用于对车辆损伤图像进行损伤分割。
根据一个实施例，所述第一类别子集中包括第一类别，所述确定权重迁移函数，包括：获取与所述第一类别对应的目标检测参数；将与所述第一类别对应的目标检测参数输入初始权重迁移函数中，得到对应的预测目标分割参数；基于所述预测目标分割参数，确定与第一类别的分割样本对应的预测目标分割结果；至少基于所述预测目标分割结果和第一类别的分割样本中标注的目标分割结果，调整所述初始权重迁移函数。
进一步地，在一个具体的实施例中，所述与所述第一类别对应的目标检测参数包括目标分类参数和目标边框参数；所述将与所述第一类别对应的目标检测参数输入初始权重迁移函数中，包括：将所述目标分类参数，和/或，目标边框参数输入所述初始权重迁移函数中。
根据一个实施例，其中所述权重迁移函数通过卷积神经网络实现。
根据一个实施例，其中所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
根据一个实施例，其中所述多个分割样本中包括目标对象为损伤对象的多个损伤分割样本。
进一步地，在一个具体的实施例中，其中所述获取同时标注有目标检测结果和目标分割结果的多个分割样本，包括：从损伤样本库中获取预定数量的损伤检测样本；基于显著性检测，从所述预定数量的损伤检测样本标注的多个边框中，提取出多个显著性区域，以使工作人员根据所述多个显著性区域，对所述预定数量中部分数量的损伤检测样本进行分割结果的标注；将标注有分割结果的部分数量的损伤检测样本，确定为所述多个损伤分割样本；以及，将所述预定数量中部分数量以外的损伤检测样本作为所述多个损伤检测样本。
根据第二方面，提供一种用于对车辆损伤图像进行部件分割的方法。该方法包括：获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第三类别子集；获取标注有目标检测结果但未标注目标 分割结果的多个部件检测样本；所述多个部件检测样本中对目标对象标注的多个类别构成第四类别子集，所述第四类别子集中包括不属于第三类别子集的类别；利用所述多个分割样本和所述多个部件检测样本，训练部件分割模型，所述部件分割模型包括用于检测部件对象的目标检测参数和用于确定所述部件对象轮廓的目标分割参数；其中所述训练部件分割模型包括：
基于所述多个分割样本中的目标检测结果和多个部件检测样本中的目标检测结果，确定与所述第三类别子集和第四类别子集构成的类别集合中各个类别对应的目标检测参数；基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述部件分割模型，用于对车辆损伤图像进行部件分割。
根据一个实施例，其中所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
根据第三方面，提供一种用于对车辆损伤图像进行损伤分割的装置，该装置包括：第一获取单元，配置为获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第一类别子集；第二获取单元，配置为获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本；所述多个损伤检测样本中对目标对象标注的多个类别构成第二类别子集，所述第二类别子集中包括不属于第一类别子集的类别；训练单元，配置为利用所述多个分割样本和所述多个损伤检测样本，训练损伤分割模型，所述损伤分割模型包括用于检测损伤对象的目标检测参数和用于确定所述损伤对象轮廓的目标分割参数；其中所训练单元具体包括：
第一确定模块，配置为基于所述多个分割样本中的目标检测结果和多个损伤检测样本中的目标检测结果，确定与所述第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数；第二确定模块，配置为基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；输入模块，配置为将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述损伤分割模型，用于对车辆损伤图像进行损伤分割。
根据第四方面，提供一种用于对车辆损伤图像进行部件分割的装置。该装置包括： 第一获取单元，配置为获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第三类别子集；第二获取单元，配置为获取标注有目标检测结果但未标注目标分割结果的多个部件检测样本；所述多个部件检测样本中对目标对象标注的多个类别构成第四类别子集，所述第四类别子集中包括不属于第三类别子集的类别；训练单元，配置为利用所述多个分割样本和所述多个部件检测样本，训练部件分割模型，所述部件分割模型包括用于检测部件对象的目标检测参数和用于确定所述部件对象轮廓的目标分割参数；其中所述训练单元具体包括：
第一确定模块，配置为基于所述多个分割样本中的目标检测结果和多个部件检测样本中的目标检测结果，确定与所述第三类别子集和第四类别子集构成的类别集合中各个类别对应的目标检测参数；第二确定模块，配置为基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；输入模块，配置为将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述部件分割模型，用于对车辆损伤图像进行部件分割。
根据第五方面，提供了一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行第一方面或第二方面所描述的方法。
根据第六方面，提供了一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现第一方面或第二方面所描述的方法。
在本说明书实施例披露的用于对车辆损伤图像进行损伤分割的方法中，可以基于与第一类别子集对应的同时标注有目标检测结果和目标分割结果的分割样本，以及与第二类别子集对应的标注有目标检测结果但未标注目标分割结果的损伤检测样本，确定与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数和目标分割参数，从而确定损伤分割模型，用于对车辆损伤图像进行损伤分割。
附图说明
为了更清楚地说明本说明书披露的多个实施例的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本说明书披露的多个实施例，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它的附图。
图1示出根据一个例子的车辆局部照片；
图2示出根据一个例子的定损客户端的界面变化示意图；
图3示出根据一个实施例的用于对车辆损伤图像进行损伤分割的方法流程图；
图4示出根据一个实施例的损伤分割样本的获取方法流程示意图；
图5示出根据一个实施例的基于显著性检测的损伤对象提取示意图；
图6示出根据一个实施例的损伤分割模型的训练方法流程图；
图7示出根据一个实施例的基于迁移学习的Mask R-CNN架构图；
图8示出根据一个实施例的用于对车辆损伤图像进行部件分割的方法流程图；
图9示出根据一个实施例的用于对车辆损伤图像进行损伤分割的装置结构图；
图10示出根据一个实施例的用于对车辆损伤图像进行部件分割的装置结构图。
具体实施方式
下面结合附图，对本说明书披露的多个实施例进行描述。
本说明书实施例披露一种用于对车辆损伤图像进行损伤分割的方法，其中损伤分割是指对车辆损伤图像中具有目标类别和明确边界的损伤对象进行区域提取，且区域提取可以表现为确定损伤对象的轮廓。
所述方法具体包括确定损伤分割模型，由此可以将车辆损伤图像输入该模型中，进而得到损伤对象的目标类别(以下简称损伤类别)和分割轮廓。下面，首先对损伤分割模型的应用场景进行介绍。
损伤分割模型可以应用于提供给用户使用的定损客户端中。根据一个例子，在事故现场，用户可以通过终端，如，手机、平板电脑等，拍摄现场照片，例如图1中示出的车辆局部图片，并将拍摄的照片上传到终端中的定损客户端，然后定损客户端可以利用损伤分割模型，确定现场照片所对应的车辆损伤信息，例如，如图2所示，可以确定出车辆损伤类别为中度刮擦，以及刮擦损伤的轮廓。进一步地，还可以给出与损伤信息对应的维修方案和相关赔偿金额，例如，维修方案为：补漆，保修赔付金额：120元。
对于损伤分割模型的训练，在一种实施方案中，可以采用传统的机器学习方法，基 于大量的人工标注的损伤分割样本进行训练。在这种方案中，工作人员通常采用笔刷等绘画工具，近似绘出损伤区域，进行损伤像素点标注。这种标注方法不精准，尤其对于大量的不连续、不规则的损伤，如刮擦，变形，撕裂等，因其损伤区域的像素点边界难以确认，而往往采用整片标注，致使标注中存在大量像素级噪声。同时，这种方案中人工标注的成本过高，因此难以实施。
基于以上观察和统计，本说明书实施例披露一种用于对车辆损伤图像进行损伤分割的方法，采用迁移学习(Transfer Learning)的方法来确定损伤分割模型，其中迁移学习可以理解为运用已有的知识来学习新的知识，其核心是找到已有知识和新知识之间的相似性。进一步地，在一种实施方式中，可以基于与损伤对象具有相似特征的非损伤对象，所对应大量分割样本，且这些分割样本通常比较容易获取，确定损伤分割模型。例如，可以采用目标对象为部件对象的分割样本，因部件的轮廓特征与损伤的轮廓特征相似，且因部件轮廓相对损伤轮廓来说较为规整，因而获取大量标注有分割结果的部件分割样本是比较容易的。
根据一个具体的实施例，首先，可以获取与车辆损伤具有相似特征的多个分割样本，以及多个损伤检测样本，其中分割样本中标注有目标检测结果(包括目标对象的类别和边框)和目标分割结果，且多个分割样本对应的多个类别构成类别集合A，而损伤检测样本中标注有目标检测结果但未标注目标分割结果，且多个损伤检测样本对应的多个类别构成类别集合B，其中类别集合B中具有不属于类别集合A的类别；然后，确定分割样本中各个类别对应的目标检测结果与目标分割结果之间的关系，再将这种关系迁移到损伤检测样本中，进而可以基于这种关系确定出类别集合B中各个类别所对应的目标分割参数，用于对车辆损伤图像中与类别集合B对应的损伤对象进行分割。下面，描述以上过程的具体实施步骤。
图3示出根据一个实施例的用于对车辆损伤图像进行损伤分割的方法流程图，所述方法的执行主体可以为具有处理能力的设备：服务器或者系统或者装置。如图3所示，该方法流程包括以下步骤：步骤S310，获取同时标注有目标检测结果和目标分割结果的多个分割样本，这些分割样本中对目标对象标注的多个类别构成第一类别子集；步骤S320，获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本，这些检测样本中对目标对象标注的多个类别构成第二类别子集，且第二类别子集中包括不属于第一类别子集的类别；步骤S330，利用所述多个分割样本和所述多个损伤检测样本，训练损伤分割模型，该模型中包括用于检测损伤对象的目标检测参数和用于确定损伤对象 轮廓的目标分割参数。
首先，在步骤S310，获取同时标注有目标检测结果和目标分割结果的多个分割样本。
具体地，其中目标检测结果包括目标对象的类别和边框，目标分割结果包括目标对象的轮廓，或者说目标对象与轮廓对应的掩码(Mask)。需要说明的是，其中目标对象为与车辆损伤具有类似或相近特征的对象，如此可以保证训练出来的损伤分割模型能够得到较好的分割效果。
在一个实施例中，目标对象可以包括车辆部件，则多个分割样本中可以包括多个部件分割样本。进一步地，根据一个具体的实施例，多个部件分割样本中可以包括基于人工标注而得到的部件分割样本。需要说明的是，因车辆损伤图像中车辆部件的轮廓较为规整，因此对车辆部件的分割结果进行人工标注的可行性较高。根据另一个具体的实施例，多个部件分割样本中还可以包括基于部件分割模型而得到的部件分割样本。在一个例子中，可以基于前述人工标注的部件分割样本训练得到部件分割模型，然后将大量的车辆损伤图片输入部件分割模型中，从而得到更多的具有部件分割结果的图像，并将这些图像作为多个部件分割样本的一部分。
在另一个实施例中，目标对象可以包括车辆损伤，则多个分割样本中可以包括多个损伤分割样本。需要说明的是，这些损伤分割样本对应的类别的数量要少于，或者说，远远少于后续将提及的损伤检测样本对应的类别的数量。进一步地，如图4所示，可以通过以下步骤获取损伤分割样本：
首先，在步骤S41，从损伤样本库中获取预定数量的损伤检测样本。
具体地，其中损伤检测样本是指标注有目标检测结果但未标注目标分割结果的样本，也就是说，损伤检测样本中标注有损伤对象的类别和损伤对象所在的边框，但是没有标注出损伤对象的分割结果。
可以理解，损伤检测样本的获取比较容易，在一个例子中，在车辆损伤图像中人工标注出损伤框和损伤类别的操作比较简单，因此可以基于人工标注获取大量的损伤检测样本。在另一个例子中，可以基于现有的损伤检测模型，以及从保险公司获取的海量车辆损伤图片，获取对应的损伤检测结果，并根据损伤检测结果确定损伤检测样本。
在一个具体的实施例中，其中预定数量可以由工作人员根据实际经验而确定。
然后，在步骤S42，基于显著性检测，从获取的预定数量的损伤检测样本标注的多个边框中，提取出多个显著性区域，以使工作人员根据多个显著性区域，对预定数量中 部分数量的损伤检测样本进行分割结果的标注。
需要说明的是，其中显著性检测(或称为视觉注意机制)是指通过智能算法模拟人的视觉特点，提取图像中的显著性区域(即人类感兴趣的区域)。通常，显著性区域具有一定的结构和纹理，色彩上具有较强的刺激，且显著性区域和周围区域具有较大的差异，基于这些简单的显著性检测原则，可以实现显著性检测。
对于损伤检测样本标注的多个边框，其中的损伤对象与边框中其他区域之间在颜色、结构和纹理上通常具有较大的反差。因此，可以通过对边框中的图像进行显著性检测，快速提取出其中的损伤对象。在一个具体的例子中，根据图5中(a)示出的车身局部图像中的损伤框，可以得到图5中(b)示出的显著性度量图，其中白色部分为检测到的损伤对象。由此，可以从预定数量的损伤检测样本的多个边框中，对应提取出多个损伤对象。
进一步地，在一个具体的实施例中，可以将基于显著性检测提取出的显著性区域，直接自动标注为对应的损伤对象的分割结果。然而，考虑到一方面，由于边框中可能包括灰尘、污渍等除损伤对象以外的显著性对象，另一方面，由于车辆损伤的种类繁多，且其中包括大量的不连续损伤，或者程度比较轻微的损伤，因此对于某些损伤检测样本，显著性检测很可能无法从中完整、准确地提取出损伤对象覆盖的区域，由此，在另一个具体的实施例中，可以将提取出的多个显著性区域提供给工作人员，以使工作人员根据多个显著性区域，对预定数量中部分数量的损伤检测样本进行分割结果的标注，例如，可以包括对显著性区域进行筛选或修正，从而得到分割结果更加精准的损伤分割样本。
以上，可以得到标注有分割结果的部分数量的损伤检测样本，接着，在步骤S43中，将得到的部分数量的损伤检测样本作为多个损伤分割样本。
可以理解的是，获取的多个分割样本，可以包括上述的多个部件分割样本，和/或，基于显著性检测得到的多个损伤分割样本。此外，在一个具体的实施例中，多个分割样本中还可以包括少量的基于人工标注而得到的损伤分割样本。在另一个具体的实施例中，多个分割样本中还可以包括其他与损伤对象具有类似特征的分割样本，例如，目标对象为植物叶子的分割样本等。
以上，可以获取同时标注有目标检测结果和目标分割结果的多个分割样本，且这些分割样本中对目标对象标注的多个类别构成第一类别子集。接着，在步骤S320，获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本。
需要说明的是，其中多个损伤检测样本中对目标对象标注的多个类别构成第二类别子集，且第二类别子集中包括不属于第一类别子集的类别。在一个实施例中，步骤S310中获取的多个分割样本包括目标对象为部件对象的部件分割样本，而损伤检测样本的目标对象为损伤对象，显然，第一类别子集中的类别为部件类别，而第二类别子集中的类别为损伤类别，损伤类别不同于部件类别。在另一个实施例中，步骤S310中获取的多个分割样本包括部分数量的损伤分割样本，由前述可知，这些损伤分割样本所对应的损伤类别是比较有限的，因此损伤分割样本对应的损伤类别的数量要少于，或者说远远少于第二类别子集中包括的损伤类别的数量。
此外，对损伤检测样本的描述，可以参见前述步骤S41中的相关描述。
根据一个具体的实施例，在获取预定数量的损伤检测样本，并得到部分数量的标注有分割结果的损伤分割样本以后，可以将剩余的(如，预定数量减去部分数量的)损伤检测样本作为本步骤中获取的多个损伤检测样本的一部分。
以上，可以在步骤S310中获取多个分割样本，以及在步骤S320中获取多个损伤检测样本。然后，在步骤S330，利用多个分割样本和多个损伤检测样本，训练损伤分割模型，此损伤分割模型包括用于检测损伤对象的目标检测参数和用于确定所述损伤对象轮廓的目标分割参数。
下面，结合图6对损伤分割模型的训练过程进行说明。如图6所示，训练损伤分割模型可以包括以下步骤：
首先，在步骤S61，基于多个分割样本中的目标检测结果和多个损伤检测样本中的目标检测结果，确定与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数。
在一个实施例中，可以采用现有的目标检测算法，例如，Faster R-CNN，R-FCN或SSD等，确定与类别集合中各个类别对应的目标检测参数。
接着，在步骤S62，基于多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射。
由于权重迁移函数表示从目标检测参数到目标分割参数的映射，在一个实施例中，分别确定出目标检测参数和目标分割参数，然后通过数学方法确定两者之间的映射关系作为权重迁移函数。具体地，首先，可以根据分割样本中的目标检测结果，确定与第一类别子集中各个类别对应的目标检测参数，以及结合目标分割结果，确定与第一类类别 子集中各个类别对应的目标分割参数；然后，可以根据与各个类别对应的目标检测参数和目标分割参数，确定权重迁移函数。
在另一个实施例中，通过训练的方式确定权重迁移函数。在该实施例中，首先确定初始权重迁移函数(即确定初步函数参数)，然后以目标检测参数为输入，以目标分割结果为标签，通过函数参数调整，训练得到权重迁移函数。下面以第一类别子集中包括的某个类别(下文称为第一类别)为例，描述训练过程的部分步骤。首先，获取与第一类别对应的目标检测参数；接着，将与第一类别对应的目标检测参数输入初始权重迁移函数中，得到对应的预测目标分割参数；然后，基于预测目标分割参数，确定与第一类别的分割样本对应的预测目标分割结果；再至少基于预测目标分割结果和第一类别的分割样本中标注的目标分割结果，调整初始权重迁移函数。如此，训练得到权重迁移函数。
在一个具体的实施例中，与第一类别对应的目标检测参数包括目标分类参数和目标边框参数，相应地，将与第一类别对应的目标检测参数输入初始权重迁移函数中可以包括：将目标分类参数，和/或，目标边框参数输入初始权重迁移函数中。在一个例子中，目标检测参数可以为目标分类参数和目标边框参数的组合。
在一个具体的实施例中，权重迁移函数可以表示为
Figure PCTCN2019097904-appb-000001
其中，ω det表示任一类别的目标检测函数；ω seg表示与ω det对应相同类别的目标分割参数，θ为与类别无关的学习参数。在初始权重迁移函数中，可以将θ设置为θ 0，具体可以随机取值。
在一个具体的实施例中，权重迁移函数可以通过卷积神经网络实现，也就是通过神经网络的神经元的组合，实现权重迁移函数的运算。相应地，权重迁移函数的训练即包括，调整和确定神经网络中神经元的运算参数和神经元之间的连接权重参数等。
在一个具体的实施例中，调整初始权重迁移函数可以包括，基于与第一类别集合中一个或者多个类别对应的预测目标分割结果和分割样本中标注的目标分割结果，确定与目标分割结果对应的损失函数，通过误差反向传播，或者梯度下降的方式，调整初始权重迁移函数。
根据一个具体的例子，首先获取与第一类别c对应的目标检测参数
Figure PCTCN2019097904-appb-000002
接着，将
Figure PCTCN2019097904-appb-000003
输入下式表示的初始权重迁移函数中：
Figure PCTCN2019097904-appb-000004
并得到与第一类别c对应的目标分割参数
Figure PCTCN2019097904-appb-000005
然后，基于
Figure PCTCN2019097904-appb-000006
确定与第一类别的分割样本对应的预测目标分割结果；再基于此预测目标分割结果和第一类别c的分割样本中标注的目标分割结果，调整初始权重迁移函数，即，将θ 0调整为θ 1。
进一步地，在对初始权重迁移函数进行调整以后，接着可以基于此次调整，用得到的分割样本以外的分割样本，进行后续调整，以确定权重迁移函数。在一个例子中，确定权重迁移函数包括确定公式(1)中θ的参数值。
需要说明的是，步骤S61和步骤S62可以同时独立进行，也可以先后独立进行，还可以将步骤S62作为步骤S61的一个分支同时进行，以节省资源开销。
以上，可以确定权重迁移函数，此函数表示从目标检测参数到目标分割参数的映射。然后，在步骤S63，将步骤S61中确定的类别集合中各个类别对应的目标检测参数输入权重迁移函数中，得到各个类别对应的目标分割参数，从而确定损伤分割模型，用于对车辆损伤图像进行损伤分割。
如此，可以得到与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标分割参数，可以用于对图像中包括的与类别集合中各个类别对应的目标对象进行分割。这也意味着，对于第二类别子集中的部分类别，在训练样本集中只存在与之对应的检测样本，而不存在与之对应的分割样本的情况下，采用本说明书实施例提供的方法，也可以得到对应的目标分割参数，用于对于这些类别对应的车辆损伤图像进行损伤分割。
下面结合一个具体的例子，对步骤S330进行进一步说明。如图7所示，其中示出的基于迁移学习的Mask R-CNN架构图，其中，假定第一类别子集为集合A，第二类别子集为集合B，A∪B表示集合A和集合B的并集，也就是前述的类别集合。
1、对于输入训练样本集中的任意一张图像(image)，首先，采用共享的卷积层(ConvNet)对全图进行特征提取，然后将得到的特征图(feature maps)送入区域生成网络(region proposal network，简称RPN)，以生成待检测区域，并通过ROI Align从特征图中提取并聚集待检测区域中的区域特征。
当此图像为与分割样本对应的图像时：
1)通过边框头部(box head)从聚集的区域特征中提取边框特征(box features)，基于边框特征和目标检测参数(box weights)，可以表示为w det，确定出预测检测结果(box predictions)，再基于预测检测结果和与此图像对应的标注检测结果(box lables in A∪B)，计算与检测结果对应的损失函数(box loss)，以调整与此图像类别所对应的 w det；
2)进而基于权重迁移函数(weight transfer function)确定与该类别对应的目标分割参数(mask weights)，可以表示为w seg，其中权重迁移函数用于表示目标检测参数和目标分割参数之间的映射关系。
3)通过分割头部(mask head)从聚集的区域特征中提取分割特征(mask features)，基于分割特征和w seg，确定预测分割结果，再基于预测分割结果和与此图像对应的标注分割结果，计算与分割结果对应的损失函数(mask loss)，进而调整权重迁移函数，也就是调整权重迁移函数中的学习参数。
由上，基于多个分割样本，可以确定出权重迁移函数。
另一方面，当此图像为与检测样本对应的图像时，对其进行的处理与上述1)中的相同，进而可以得到与此图像的类别对应的目标检测参数。
基于以上处理，可以基于与类别集合A对应的多个分割样本，确定权重迁移函数，同时，基于类别集合A对应的多个分割样本和类别集合B对应的多个损伤检测样本，确定与A∪B中各个类别对应的目标检测参数。
2、将与A∪B中各个类别对应的目标检测参数输入权重迁移函数中，得到各个类别对应的目标分割参数，从而确定损伤分割模型，用于对与A∪B中的损伤类别对应的车辆损伤图像进行损伤分割。
综上可知，采用本说明书实施例提供的用于损伤分割的方法，可以基于与第一类别子集对应的同时标注有目标检测结果和目标分割结果的分割样本，以及与第二类别子集对应的标注有目标检测结果但未标注目标分割结果的损伤检测样本，确定与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数和目标分割参数，从而确定损伤分割模型，用于对车辆损伤图像进行损伤分割。
根据另一方面的实施例，还提供一种用于对车辆损伤图像进行部件分割的方法。图8示出根据一个实施例的用于对车辆损伤图像进行损伤分割的方法流程图，所述方法的执行主体可以为具有处理能力的设备：服务器或者系统或者装置。如图8所示，该方法流程包括以下步骤：步骤S810，获取同时标注有目标检测结果和目标分割结果的多个分割样本，这些分割样本中对目标对象标注的多个类别构成第三类别子集；步骤S820，获取标注有目标检测结果但未标注目标分割结果的多个部件检测样本，这些检测样本中对目标对象标注的多个类别构成第四类别子集，且第四类别子集中包括不属于第三类别 子集的类别；步骤S830，利用多个分割样本和多个部件检测样本，训练部件分割模型，该模型中包括用于检测部件对象的目标检测参数和用于确定所述部件对象轮廓的目标分割参数。
首先，在步骤S810，获取同时标注有目标检测结果和目标分割结果的多个分割样本，这些分割样本中对目标对象标注的多个类别构成第三类别子集。
在一个实施例中，多个分割样本中可以包括与部件具有类似轮廓特征的分割样本，如多种植物的树叶分割样本等。
在另一个实施例中，多个分割样本中包括目标对象为部件对象的多个部件分割样本。
需要说明的是，因车辆的部件种类繁多，获取到的多个部件分割样本所对应的类别，通常只是车辆部件中的一部分，在一个例子中，可能占所有车辆部件类别的1/3或1/4。而对于其他类别的车辆部件，与之对应的部件检测样本较易获取，而标注有分割结果的部件分割样本较难获取。比如说，对于有些车辆部件，在车辆事故中，其发生损伤的可能性较小，因此与其相关的车辆损伤图片有限，此外，因标注分割结果的成本较高，因此对这类部件的分割结果进行人工标注的可能性较少，进而难以获取这些部件的分割样本。
此外，需要说明的是，对步骤S810-步骤S830的描述，还可以参见前述对步骤S310-步骤S330的描述，在此不作赘述。
综上可知，采用本说明书实施例提供的用于部件分割的方法，可以基于与第三类别子集对应的同时标注有目标检测结果和目标分割结果的分割样本，以及与第四类别子集对应的标注有目标检测结果但未标注目标分割结果的部件检测样本，确定与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数和目标分割参数，从而确定部件分割模型，用于对车辆损伤图像进行部件分割。
根据再一方面的实施例，还提供一种用于损伤分割的装置。图9示出根据一个实施例的用于对车辆损伤图像进行损伤分割的装置结构图。如图9所示，该装置包括：
第一获取单元910，配置为获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第一类别子集；
第二获取单元920，配置为获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本；所述多个损伤检测样本中对目标对象标注的多个类别构成第二类别子集， 所述第二类别子集中包括不属于第一类别子集的类别；
训练单元930，配置为利用所述多个分割样本和所述多个损伤检测样本，训练损伤分割模型，所述损伤分割模型包括用于检测损伤对象的目标检测参数和用于确定所述损伤对象轮廓的目标分割参数；其中所训练单元930具体包括：
第一确定模块931，配置为基于所述多个分割样本中的目标检测结果和多个损伤检测样本中的目标检测结果，确定与所述第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数；
第二确定模块932，配置为基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；
输入模块933，配置为将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述损伤分割模型，用于对车辆损伤图像进行损伤分割。
根据一个实施例，所述第一类别子集中包括第一类别，所述第二确定模块932具体配置为：
获取与所述第一类别对应的目标检测参数；
将与所述第一类别对应的目标检测参数输入初始权重迁移函数中，得到对应的预测目标分割参数；
基于所述预测目标分割参数，确定与第一类别的分割样本对应的预测目标分割结果；
至少基于所述预测目标分割结果和第一类别的分割样本中标注的目标分割结果，调整所述初始权重迁移函数。
进一步地，在一个具体的实施例中，所述与所述第一类别对应的目标检测参数包括目标分类参数和目标边框参数；所述输入模块933具体配置为：
将所述目标分类参数，和/或，目标边框参数输入所述初始权重迁移函数中。
根据一个实施例，其中所述权重迁移函数通过卷积神经网络实现。
根据一个实施例，所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
根据一个实施例，所述多个分割样本中包括目标对象为损伤对象的多个损伤分割样本。
进一步地，在一个具体的实施例中，所述第一获取单元910具体配置为：
从损伤样本库中获取预定数量的损伤检测样本；
基于显著性检测，从所述预定数量的损伤检测样本标注的多个边框中，提取出多个显著性区域，以使工作人员根据所述多个显著性区域，对所述预定数量中部分数量的损伤检测样本进行分割结果的标注；
将标注有分割结果的部分数量的损伤检测样本，确定为所述多个损伤分割样本；以及，将所述预定数量中部分数量以外的损伤检测样本作为所述多个损伤检测样本。
综上可知，采用本说明书实施例提供的用于损伤分割的装置，可以基于与第一类别子集对应的同时标注有目标检测结果和目标分割结果的分割样本，以及与第二类别子集对应的标注有目标检测结果但未标注目标分割结果的损伤检测样本，确定与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数和目标分割参数，从而确定损伤分割模型，用于对车辆损伤图像进行损伤分割。
根据再一方面的实施例，还提供一种用于部件分割的装置。图10示出根据一个实施例的用于对车辆损伤图像进行部件分割的装置结构图。如图10所示，该装置包括：
第一获取单元1010，配置为获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第三类别子集；
第二获取单元1020，配置为获取标注有目标检测结果但未标注目标分割结果的多个部件检测样本；所述多个部件检测样本中对目标对象标注的多个类别构成第四类别子集，所述第四类别子集中包括不属于第三类别子集的类别；
训练单元1030，配置为利用所述多个分割样本和所述多个部件检测样本，训练部件分割模型，所述部件分割模型包括用于检测部件对象的目标检测参数和用于确定所述部件对象轮廓的目标分割参数；其中所述训练单元1030具体包括：
第一确定模块1031，配置为基于所述多个分割样本中的目标检测结果和多个部件检测样本中的目标检测结果，确定与所述第三类别子集和第四类别子集构成的类别集合中各个类别对应的目标检测参数；
第二确定模块1032，配置为基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；
输入模块1033，配置为将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述部件分割模型，用于对车辆损伤图像进行部件分割。
在一个实施例中，所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
综上可知，采用本说明书实施例提供的用于部件分割的装置，可以基于与第三类别子集对应的同时标注有目标检测结果和目标分割结果的分割样本，以及与第四类别子集对应的标注有目标检测结果但未标注目标分割结果的部件检测样本，确定与第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数和目标分割参数，从而确定部件分割模型，用于对车辆损伤图像进行部件分割。
如上，根据又一方面的实施例，还提供一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算机中执行时，令计算机执行结合图3、图4、图6或图8所描述的方法。
根据又一方面的实施例，还提供一种计算设备，包括存储器和处理器，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现结合图3、图4、图6或图8所描述的方法。
本领域技术人员应该可以意识到，在上述一个或多个示例中，本说明书披露的多个实施例所描述的功能可以用硬件、软件、固件或它们的任意组合来实现。当使用软件实现时，可以将这些功能存储在计算机可读介质中或者作为计算机可读介质上的一个或多个指令或代码进行传输。
以上所述的具体实施方式，对本说明书披露的多个实施例的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本说明书披露的多个实施例的具体实施方式而已，并不用于限定本说明书披露的多个实施例的保护范围，凡在本说明书披露的多个实施例的技术方案的基础之上，所做的任何修改、等同替换、改进等，均应包括在本说明书披露的多个实施例的保护范围之内。

Claims (20)

  1. 一种用于对车辆损伤图像进行损伤分割的方法，所述方法包括：
    获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第一类别子集；
    获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本；所述多个损伤检测样本中对目标对象标注的多个类别构成第二类别子集，所述第二类别子集中包括不属于第一类别子集的类别；
    利用所述多个分割样本和所述多个损伤检测样本，训练损伤分割模型，所述损伤分割模型包括用于检测损伤对象的目标检测参数和用于确定所述损伤对象轮廓的目标分割参数；其中所述训练损伤分割模型包括：
    基于所述多个分割样本中的目标检测结果和多个损伤检测样本中的目标检测结果，确定与所述第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数；
    基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；
    将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述损伤分割模型，用于对车辆损伤图像进行损伤分割。
  2. 根据权利要求1所述的方法，其中，所述第一类别子集中包括第一类别，所述确定权重迁移函数，包括：
    获取与所述第一类别对应的目标检测参数；
    将与所述第一类别对应的目标检测参数输入初始权重迁移函数中，得到对应的预测目标分割参数；
    基于所述预测目标分割参数，确定与第一类别的分割样本对应的预测目标分割结果；
    至少基于所述预测目标分割结果和第一类别的分割样本中标注的目标分割结果，调整所述初始权重迁移函数。
  3. 根据权利要求2所述的方法，其中，所述与所述第一类别对应的目标检测参数包括目标分类参数和目标边框参数；所述将与所述第一类别对应的目标检测参数输入初始权重迁移函数中，包括：
    将所述目标分类参数，和/或，目标边框参数输入所述初始权重迁移函数中。
  4. 根据权利要求1所述的方法，其中所述权重迁移函数通过卷积神经网络实现。
  5. 根据权利要求1所述的方法，其中，所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
  6. 根据权利要求1所述的方法，其中，所述多个分割样本中包括目标对象为损伤对象的多个损伤分割样本。
  7. 根据权利要求6所述的方法，其中，所述获取同时标注有目标检测结果和目标分割结果的多个分割样本，包括：
    从损伤样本库中获取预定数量的损伤检测样本；
    基于显著性检测，从所述预定数量的损伤检测样本标注的多个边框中，提取出多个显著性区域，以使工作人员根据所述多个显著性区域，对所述预定数量中部分数量的损伤检测样本进行分割结果的标注；
    将标注有分割结果的部分数量的损伤检测样本，确定为所述多个损伤分割样本；以及，将所述预定数量中部分数量以外的损伤检测样本作为所述多个损伤检测样本。
  8. 一种用于对车辆损伤图像进行部件分割的方法，所述方法包括：
    获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第三类别子集；
    获取标注有目标检测结果但未标注目标分割结果的多个部件检测样本；所述多个部件检测样本中对目标对象标注的多个类别构成第四类别子集，所述第四类别子集中包括不属于第三类别子集的类别；
    利用所述多个分割样本和所述多个部件检测样本，训练部件分割模型，所述部件分割模型包括用于检测部件对象的目标检测参数和用于确定所述部件对象轮廓的目标分割参数；其中所述训练部件分割模型包括：
    基于所述多个分割样本中的目标检测结果和多个部件检测样本中的目标检测结果，确定与所述第三类别子集和第四类别子集构成的类别集合中各个类别对应的目标检测参数；
    基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；
    将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述部件分割模型，用于对车辆损伤图像进行部件分割。
  9. 根据权利要求8所述的方法，其中，所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
  10. 一种用于对车辆损伤图像进行损伤分割的装置，所述装置包括：
    第一获取单元，配置为获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第一类别子集；
    第二获取单元，配置为获取标注有目标检测结果但未标注目标分割结果的多个损伤检测样本；所述多个损伤检测样本中对目标对象标注的多个类别构成第二类别子集，所述第二类别子集中包括不属于第一类别子集的类别；
    训练单元，配置为利用所述多个分割样本和所述多个损伤检测样本，训练损伤分割模型，所述损伤分割模型包括用于检测损伤对象的目标检测参数和用于确定所述损伤对象轮廓的目标分割参数；其中所训练单元具体包括：
    第一确定模块，配置为基于所述多个分割样本中的目标检测结果和多个损伤检测样本中的目标检测结果，确定与所述第一类别子集和第二类别子集构成的类别集合中各个类别对应的目标检测参数；
    第二确定模块，配置为基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；
    输入模块，配置为将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述损伤分割模型，用于对车辆损伤图像进行损伤分割。
  11. 根据权利要求10所述的装置，其中，所述第一类别子集中包括第一类别，所述第二确定模块具体配置为：
    获取与所述第一类别对应的目标检测参数；
    将与所述第一类别对应的目标检测参数输入初始权重迁移函数中，得到对应的预测目标分割参数；
    基于所述预测目标分割参数，确定与第一类别的分割样本对应的预测目标分割结果；
    至少基于所述预测目标分割结果和第一类别的分割样本中标注的目标分割结果，调整所述初始权重迁移函数。
  12. 根据权利要求11所述的装置，其中，所述与所述第一类别对应的目标检测参数包括目标分类参数和目标边框参数；所述输入模块具体配置为：
    将所述目标分类参数，和/或，目标边框参数输入所述初始权重迁移函数中。
  13. 根据权利要求10所述的装置，其中所述权重迁移函数通过卷积神经网络实现。
  14. 根据权利要求10所述的装置，其中，所述多个分割样本中包括目标对象为部 件对象的多个部件分割样本。
  15. 根据权利要求10所述的装置，其中，所述多个分割样本中包括目标对象为损伤对象的多个损伤分割样本。
  16. 根据权利要求15所述的装置，其中，所述第一获取单元具体配置为：
    从损伤样本库中获取预定数量的损伤检测样本；
    基于显著性检测，从所述预定数量的损伤检测样本标注的多个边框中，提取出多个显著性区域，以使工作人员根据所述多个显著性区域，对所述预定数量中部分数量的损伤检测样本进行分割结果的标注；
    将标注有分割结果的部分数量的损伤检测样本，确定为所述多个损伤分割样本；以及，将所述预定数量中部分数量以外的损伤检测样本作为所述多个损伤检测样本。
  17. 一种用于对车辆损伤图像进行部件分割的装置，所述装置包括：
    第一获取单元，配置为获取同时标注有目标检测结果和目标分割结果的多个分割样本，所述目标检测结果包括目标对象的类别和边框，所述目标分割结果包括目标对象的轮廓；所述多个分割样本中对目标对象标注的多个类别构成第三类别子集；
    第二获取单元，配置为获取标注有目标检测结果但未标注目标分割结果的多个部件检测样本；所述多个部件检测样本中对目标对象标注的多个类别构成第四类别子集，所述第四类别子集中包括不属于第三类别子集的类别；
    训练单元，配置为利用所述多个分割样本和所述多个部件检测样本，训练部件分割模型，所述部件分割模型包括用于检测部件对象的目标检测参数和用于确定所述部件对象轮廓的目标分割参数；其中所述训练单元具体包括：
    第一确定模块，配置为基于所述多个分割样本中的目标检测结果和多个部件检测样本中的目标检测结果，确定与所述第三类别子集和第四类别子集构成的类别集合中各个类别对应的目标检测参数；
    第二确定模块，配置为基于所述多个分割样本中的目标检测结果和目标分割结果，确定权重迁移函数，所述权重迁移函数表示从目标检测参数到目标分割参数的映射；
    输入模块，配置为将所述各个类别对应的目标检测参数输入所述权重迁移函数中，得到各个类别对应的目标分割参数，从而确定所述部件分割模型，用于对车辆损伤图像进行部件分割。
  18. 根据权利要求17所述的装置，其中，所述多个分割样本中包括目标对象为部件对象的多个部件分割样本。
  19. 一种计算机可读存储介质，其上存储有计算机程序，当所述计算机程序在计算 机中执行时，令计算机执行权利要求1-9中任一项的所述的方法。
  20. 一种计算设备，包括存储器和处理器，其特征在于，所述存储器中存储有可执行代码，所述处理器执行所述可执行代码时，实现权利要求1-9中任一项所述的方法。
PCT/CN2019/097904 2018-09-10 2019-07-26 用于对车辆损伤图像进行损伤分割的方法及装置 Ceased WO2020052352A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
EP19858832.9A EP3852061B1 (en) 2018-09-10 2019-07-26 Method and device for damage segmentation of vehicle damage image
SG11202100117XA SG11202100117XA (en) 2018-09-10 2019-07-26 Method and apparatus for performing damage segmentation on vehicle damage image
US17/154,361 US11042978B2 (en) 2018-09-10 2021-01-21 Method and apparatus for performing damage segmentation on vehicle damage image

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811052994.6A CN110570435B (zh) 2018-09-10 2018-09-10 用于对车辆损伤图像进行损伤分割的方法及装置
CN201811052994.6 2018-09-10

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/154,361 Continuation US11042978B2 (en) 2018-09-10 2021-01-21 Method and apparatus for performing damage segmentation on vehicle damage image

Publications (1)

Publication Number Publication Date
WO2020052352A1 true WO2020052352A1 (zh) 2020-03-19

Family

ID=68772408

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/097904 Ceased WO2020052352A1 (zh) 2018-09-10 2019-07-26 用于对车辆损伤图像进行损伤分割的方法及装置

Country Status (6)

Country Link
US (1) US11042978B2 (zh)
EP (1) EP3852061B1 (zh)
CN (1) CN110570435B (zh)
SG (1) SG11202100117XA (zh)
TW (1) TWI709085B (zh)
WO (1) WO2020052352A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112183541A (zh) * 2020-09-17 2021-01-05 中山大学肿瘤防治中心 一种轮廓提取方法及装置、电子设备、存储介质
CN112365453A (zh) * 2020-10-26 2021-02-12 上海东普信息科技有限公司 快递货物堆放的检测方法、装置、设备和存储介质
CN113450362A (zh) * 2021-06-02 2021-09-28 北京迈格威科技有限公司 一种视频目标分割方法、装置和电子系统
CN114519793A (zh) * 2020-11-16 2022-05-20 顺丰科技有限公司 目标物体检测方法、装置、电子设备及存储介质

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359676A (zh) * 2018-10-08 2019-02-19 百度在线网络技术（北京）有限公司 用于生成车辆损伤信息的方法和装置
CN111091492B (zh) * 2019-12-23 2020-09-04 韶鼎人工智能科技有限公司 一种基于卷积神经网络的人脸图像光照迁移方法
WO2021136939A1 (en) 2020-01-03 2021-07-08 Tractable Ltd Method of determining repair operations for a damaged vehicle
CN112192318B (zh) * 2020-09-28 2022-07-19 北京天泽智云科技有限公司 机加工刀具状态监控方法和系统
US12211299B2 (en) * 2021-05-10 2025-01-28 Ccc Intelligent Solutions Inc. Methods and systems of utilizing image processing systems to measure objects
CN113763389B (zh) * 2021-08-24 2022-06-14 深圳前海爱客风信息技术有限公司 一种基于多主体检测分割的图像识别方法
US12002192B2 (en) * 2021-11-16 2024-06-04 Solera Holdings, Llc Transfer of damage markers from images to 3D vehicle models for damage assessment
CN115115611B (zh) 2022-07-21 2023-04-07 明觉科技（北京）有限公司 车辆损伤识别方法、装置、电子设备和存储介质
CN116523865A (zh) * 2023-04-25 2023-08-01 大连理工大学 一种工件表面粗糙度预测方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017176304A1 (en) * 2016-04-06 2017-10-12 American International Group, Inc. Automatic assessment of damage and repair costs in vehicles
CN107392218A (zh) * 2017-04-11 2017-11-24 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置及电子设备
CN107977895A (zh) * 2017-12-08 2018-05-01 北京小米移动软件有限公司 车辆损毁信息的确定方法、装置及用户设备
CN108446618A (zh) * 2018-03-09 2018-08-24 平安科技（深圳）有限公司 车辆定损方法、装置、电子设备及存储介质

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11017250B2 (en) * 2010-06-07 2021-05-25 Affectiva, Inc. Vehicle manipulation using convolutional image processing
RU2629875C1 (ru) * 2016-04-04 2017-09-04 Общество С Ограниченной Ответственностью "Яндекс" Способы и системы прогнозирования условий вождения
US11144889B2 (en) * 2016-04-06 2021-10-12 American International Group, Inc. Automatic assessment of damage and repair costs in vehicles
CN107292394A (zh) * 2016-04-11 2017-10-24 富泰华工业（深圳）有限公司 车辆损伤定价系统及方法
US10445576B2 (en) * 2016-09-23 2019-10-15 Cox Automotive, Inc. Automated vehicle recognition systems
US10546242B2 (en) * 2017-03-03 2020-01-28 General Electric Company Image analysis neural network systems
WO2018187632A1 (en) * 2017-04-05 2018-10-11 Carnegie Mellon University Deep learning methods for estimating density and/or flow of objects, and related methods and software
CN107194398B (zh) * 2017-05-10 2018-09-25 平安科技（深圳）有限公司 车损部位的识别方法及系统
US20190095877A1 (en) * 2017-09-26 2019-03-28 Panton, Inc. Image recognition system for rental vehicle damage detection and management
CN108460328A (zh) * 2018-01-15 2018-08-28 浙江工业大学 一种基于多任务卷积神经网络的套牌车检测方法
CN108414549A (zh) * 2018-03-07 2018-08-17 中国科学院高能物理研究所 一种用于汽车定损的射线检测方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017176304A1 (en) * 2016-04-06 2017-10-12 American International Group, Inc. Automatic assessment of damage and repair costs in vehicles
CN107392218A (zh) * 2017-04-11 2017-11-24 阿里巴巴集团控股有限公司 一种基于图像的车辆定损方法、装置及电子设备
CN107977895A (zh) * 2017-12-08 2018-05-01 北京小米移动软件有限公司 车辆损毁信息的确定方法、装置及用户设备
CN108446618A (zh) * 2018-03-09 2018-08-24 平安科技（深圳）有限公司 车辆定损方法、装置、电子设备及存储介质

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3852061A4 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112183541A (zh) * 2020-09-17 2021-01-05 中山大学肿瘤防治中心 一种轮廓提取方法及装置、电子设备、存储介质
CN112183541B (zh) * 2020-09-17 2023-04-18 中山大学肿瘤防治中心（中山大学附属肿瘤医院、中山大学肿瘤研究所） 一种轮廓提取方法及装置、电子设备、存储介质
CN112365453A (zh) * 2020-10-26 2021-02-12 上海东普信息科技有限公司 快递货物堆放的检测方法、装置、设备和存储介质
CN114519793A (zh) * 2020-11-16 2022-05-20 顺丰科技有限公司 目标物体检测方法、装置、电子设备及存储介质
CN113450362A (zh) * 2021-06-02 2021-09-28 北京迈格威科技有限公司 一种视频目标分割方法、装置和电子系统

Also Published As

Publication number Publication date
EP3852061A1 (en) 2021-07-21
CN110570435A (zh) 2019-12-13
EP3852061A4 (en) 2022-06-08
US11042978B2 (en) 2021-06-22
US20210150691A1 (en) 2021-05-20
EP3852061B1 (en) 2024-06-26
TWI709085B (zh) 2020-11-01
TW202011267A (zh) 2020-03-16
SG11202100117XA (en) 2021-02-25
CN110570435B (zh) 2020-06-26

Similar Documents

Publication Publication Date Title
WO2020052352A1 (zh) 用于对车辆损伤图像进行损伤分割的方法及装置
CN110705405B (zh) 目标标注的方法及装置
CN111862296B (zh) 三维重建方法及装置、系统、模型训练方法、存储介质
TWI716012B (zh) 樣本標註方法、裝置、儲存媒體和計算設備、損傷類別的識別方法及裝置
CN111340749B (zh) 图像质量的检测方法、装置、设备及存储介质
WO2018108129A1 (zh) 用于识别物体类别的方法及装置、电子设备
CN108229475B (zh) 车辆跟踪方法、系统、计算机设备及可读存储介质
CN110930427B (zh) 一种基于语义轮廓信息的图像分割方法、设备和存储介质
CN111639629B (zh) 一种基于图像处理的猪只体重测量方法、装置及存储介质
CN109711268B (zh) 一种人脸图像筛选方法及设备
WO2020238256A1 (zh) 基于弱分割的损伤检测方法及装置
WO2017120796A1 (zh) 路面病害的检测方法及其装置、电子设备
JP2017522647A (ja) 背景追跡を介する物体追跡及びセグメント化の方法及び装置
CN109993021A (zh) 人脸正脸检测方法、装置及电子设备
CN113706472A (zh) 公路路面病害检测方法、装置、设备及存储介质
CN116758421A (zh) 一种基于弱监督学习的遥感图像有向目标检测方法
WO2014205787A1 (zh) 一种基于混合图像模板的车辆检测方法
CN111754474A (zh) 一种基于图像清晰度的能见度识别的方法及设备
CN110826564A (zh) 一种复杂场景图像中的小目标语义分割方法及系统
CN110991434A (zh) 自助终端证件识别方法及装置
CN110738656B (zh) 一种证件照片的清晰度评估方法、存储介质及处理器
Zhang et al. A combined approach to single-camera-based lane detection in driverless navigation
CN114998953B (zh) 人脸关键点检测方法及装置
CN118781076A (zh) 一种基于深度卷积自编码器的艺术品缺陷检测方法
CN112115976B (zh) 模型训练方法、模型训练装置、存储介质和电子设备

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: 19858832

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2019858832

Country of ref document: EP

Effective date: 20210412