WO2020209487A1 - Dispositif de reconnaissance de lieu basé sur un réseau neuronal artificiel et son dispositif d'apprentissage - Google Patents

Dispositif de reconnaissance de lieu basé sur un réseau neuronal artificiel et son dispositif d'apprentissage Download PDF

Info

Publication number
WO2020209487A1
WO2020209487A1 PCT/KR2020/001018 KR2020001018W WO2020209487A1 WO 2020209487 A1 WO2020209487 A1 WO 2020209487A1 KR 2020001018 W KR2020001018 W KR 2020001018W WO 2020209487 A1 WO2020209487 A1 WO 2020209487A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature value
place
correlation
neural network
artificial neural
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/KR2020/001018
Other languages
English (en)
Korean (ko)
Inventor
김은태
성홍제
현준혁
이수현
우수한
장현배
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.)
University Industry Foundation UIF of Yonsei University
Original Assignee
University Industry Foundation UIF of Yonsei University
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 University Industry Foundation UIF of Yonsei University filed Critical University Industry Foundation UIF of Yonsei University
Publication of WO2020209487A1 publication Critical patent/WO2020209487A1/fr
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
    • 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
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/48—Extraction of image or video features by mapping characteristic values of the pattern into a parameter space, e.g. Hough transformation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image

Definitions

  • the present invention relates to a place recognition apparatus, and more particularly, to a place recognition apparatus based on an artificial neural network and a learning apparatus thereof.
  • This study is related to a research on the source technology for understanding the semantic context based on deep learning of a research project conducted with the support of the National Research Foundation of Korea with the funding of the Ministry of Science, ICT and Future Planning (Government).
  • the method of deriving the correlation coefficient of the conventional statistics uses a method of expressing the relationship between a specific object and a place numerically by statistically analyzing all the correct answers when there is a correct answer for which object and place appear for all images. .
  • the present invention is an artificial neural network-based place recognition device and a learning device thereof to learn a correlation feature value representing a correlation according to a combination of a plurality of objects and an object and place that can be selected from a plurality of places, or
  • the purpose of this is to improve image-unit location recognition performance by including an artificial neural network-based correlation feature value management unit that manages information on the already learned correlation feature values.
  • Still other objects, not specified, of the present invention may be additionally considered within the range that can be easily deduced from the following detailed description and effects thereof.
  • an artificial neural network-based place recognition apparatus includes an object feature value extraction unit for extracting an object feature value from an input image, a plurality of predetermined objects and a plurality of places.
  • An artificial neural network-based correlation feature value management unit that learns a correlation feature value representing a correlation according to a combination of an object and a place that can be selected from, or manages information on the already learned correlation feature value, and the object feature value
  • a place feature value generator that generates a place feature value for the input image according to the correlated feature value, and a recognizer that recognizes a place of the input image using the place feature value.
  • the artificial neural network-based correlation feature value management unit is a combination of a plurality of predetermined objects and an object that can be selected from a plurality of places among training images used to learn the artificial neural network.
  • the correlation feature value is determined in consideration of the extent to which there are training images related to the variable.
  • the correlation feature value is a current correlation feature value in consideration of a location feature value generated based on a current correlation feature value in a process in which the artificial neural network-based correlation feature value manager performs learning using the training images. It is obtained through the process of updating the feature value to the correlated feature value.
  • the correlation feature value is a matrix representing a correlation index indicating a correlation between an object and a place included in each combination variable among the training images.
  • the recognition unit recognizes the location using the location feature value and a predetermined connection weight to recognize the location.
  • An artificial neural network-based location recognition apparatus includes an object feature value extractor for extracting an object feature value from an input image, an object that can be selected from a plurality of predetermined objects and a plurality of places
  • a correlation feature value management unit based on an artificial neural network that learns a correlation feature value representing a correlation according to a combination of a and a place, or manages information on an already learned correlation feature value, according to the object feature value and the correlation feature value
  • a correlation-based feature value generator including a place feature value generator that generates a first place feature value for the input image, and a place feature value extractor that extracts a second place feature value for recognizing a place related to the input image
  • a recognition unit determining a place related to the input image by using the first place feature value and the second place feature value.
  • the object feature value extracting unit extracts an object feature value by applying the input image to a pre-learned first neural network to recognize an object from the input image, and the location feature value extracting unit is included in the input image.
  • a place feature value is extracted by applying the input image to a pre-trained second neural network.
  • the artificial neural network-based correlation feature value management unit is a combination of a plurality of predetermined objects and an object that can be selected from a plurality of places among training images used to learn the artificial neural network.
  • the correlation feature value is determined in consideration of the extent to which there are training images related to the variable.
  • the correlation feature value is a current correlation feature value in consideration of a location feature value generated based on a current correlation feature value in a process in which the artificial neural network-based correlation feature value manager performs learning using the training images. It is obtained through the process of updating the feature value to the correlated feature value.
  • the correlation feature value is a matrix representing a correlation index indicating a correlation between an object and a place included in each combination variable among the training images.
  • the recognition unit generates a synthesized place feature value by synthesizing the first place feature value and the second place feature value according to a predetermined criterion, and a predetermined connection to recognize the synthesized place feature value and the place Recognize the place using weights.
  • An artificial neural network-based learning apparatus includes a processor and a memory that perform signal processing for generating a correlation feature value based on the artificial neural network, and the processor includes an object feature value for a training image. Is extracted, and a correlation feature value for an object and a place is generated using the object feature value and the connection weight, and the connection weight is updated through error backpropagation.
  • the extraction of the object feature values for the training image includes receiving the training image and extracting the object feature values for the training image from the training image using a convolutional neural network.
  • the correlation feature value represents a correlation between a plurality of predetermined objects and an object selected from a plurality of places and each combination of a place.
  • the processor calculates an error by comparing the generated location feature value with a previously stored location feature value for the training image, and performs signal processing for transferring the error to the correlation feature value generator.
  • a correlation feature value indicating a correlation according to a combination of an object and a place that can be selected from a plurality of predetermined objects and a plurality of places is learned, or
  • a correlation feature value management unit based on an artificial neural network that manages information on the learned correlation feature value may be included to improve the performance of image unit location recognition.
  • FIG. 1 is a block diagram illustrating an apparatus for recognizing a place based on an artificial neural network according to an embodiment of the present invention.
  • FIG. 2 is a block diagram showing an apparatus for recognizing a place based on an artificial neural network according to another embodiment of the present invention.
  • FIG. 3 is a block diagram showing an artificial neural network-based learning apparatus according to an embodiment of the present invention.
  • FIG. 4 is a diagram illustrating an exemplary structure including a convolutional neural network of a place recognition apparatus based on an artificial neural network according to an embodiment of the present invention.
  • FIG. 5 is a diagram illustrating a process of calculating a location feature value vector of a learning device based on an artificial neural network according to an embodiment of the present invention.
  • 6A and 6B are graphs showing a correlation weight of an object and a correlation weight of a place according to an analysis of a place recognition apparatus based on an artificial neural network according to an embodiment of the present invention.
  • FIG. 7 is a diagram showing an exemplary structure including a convolutional neural network of a place recognition apparatus based on an artificial neural network according to another embodiment of the present invention.
  • the present invention relates to an artificial neural network-based location recognition device and a learning device thereof.
  • FIG. 1 is a block diagram illustrating an apparatus for recognizing a place based on an artificial neural network according to an embodiment of the present invention.
  • an artificial neural network-based place recognition apparatus 1 includes an object feature value extraction unit 11, a correlation feature value management unit 12, and a place feature value generation unit 13 , And a recognition unit 14.
  • the apparatus 1 for recognizing place based on an artificial neural network is an apparatus for recognizing place information from object information. Specifically, a deep learning algorithm that learns the object-to-place correlation feature value matrix using the object information extracted from the object classifier and the place information extracted from the place classifier in a weak map method, and a place using the same.
  • DNN Convolutional Neural Network
  • DNN is a type of deep neural network (DNN), one or several convolutional layers, a pooling layer, and a fully connected layer. ) Is a neural network.
  • CNN has a structure suitable for learning 2D data, and can be trained through a backpropagation algorithm. It is one of the representative models of DNN that is widely used in various application fields such as object classification and object detection in images.
  • the object feature value extraction unit 11 extracts object feature values from the input image.
  • information related to a plurality of objects included in an analysis target image of a place to be recognized is extracted as an object feature value, and the above is performed using a convolutional neural network (CNN) previously learned in the object recognition dataset.
  • CNN convolutional neural network
  • the object recognition data set is a set of areas in the analysis target image that can detect objects related to the location to be recognized.
  • the correlation feature value management unit 12 learns a correlation feature value representing a correlation according to a combination of a plurality of objects and an object that can be selected from a plurality of places based on an artificial neural network, or a correlation feature value that has already been learned. Manage information on correlation feature values.
  • the correlation feature value management unit 12 is a training image related to a combination variable that combines a plurality of predetermined objects and an object and a place that can be selected from a plurality of places among training images used for learning the artificial neural network.
  • the correlation feature value is determined in consideration of the extent to which they exist.
  • the correlation feature value is the current correlation feature value in consideration of the location feature value generated based on the current correlation feature value in the process of the artificial neural network-based correlation feature value management unit performing learning using the training images. It is obtained through the process of updating to the correlation feature value.
  • the correlation feature value is a matrix representing a correlation index representing a correlation between an object and a place included in each combination variable among the training images.
  • the place feature value generator 13 generates a place feature value for the input image according to the object feature value and the correlation feature value.
  • place information of the analysis target image is extracted as a place feature value by using an image unit place recognition dataset from an analysis target image including a place to be recognized.
  • the recognition unit 14 recognizes the location of the input image using the location feature value.
  • the recognition unit recognizes the location using the location feature value and a predetermined connection weight to recognize the location.
  • FIG. 2 is a block diagram showing an apparatus for recognizing a place based on an artificial neural network according to another embodiment of the present invention.
  • an artificial neural network-based place recognition apparatus 2 includes an object feature value extraction unit 21, a correlation feature value management unit 22, and a correlation-based feature value generation unit ( 23), a location feature value extraction unit 24, and a recognition unit 25.
  • the object feature value extracting unit 21 extracts object feature values from the input image.
  • the object feature value extraction unit 21 extracts an object feature value by applying the input image to a pre-learned first neural network, and the location feature value extraction unit 24 performs the input.
  • a place feature value is extracted by applying the input image to a pre-learned second neural network.
  • the correlation feature value management unit 22 learns a correlation feature value representing a correlation according to a combination of a plurality of objects and an object that can be selected from a plurality of places based on an artificial neural network, or a correlation feature value that has already been learned. Manage information on correlation feature values.
  • the correlation feature value management unit 22 is a training image related to a combination variable that combines a plurality of predetermined objects and an object and a place that can be selected from a plurality of places among training images used for learning the artificial neural network.
  • the correlation feature value is determined in consideration of the extent to which they exist.
  • the correlation feature value is the current correlation feature value in consideration of the location feature value generated based on the current correlation feature value in the process of the artificial neural network-based correlation feature value management unit performing learning using the training images. It is obtained through the process of updating to the correlation feature value.
  • the correlation feature value is a matrix representing a correlation index representing a correlation between an object and a place included in each combination variable among the training images.
  • the correlation-based feature value generator 23 includes a place feature value generator that generates a first place feature value for the input image according to the object feature value and the correlation feature value.
  • the place feature value extracting unit 24 extracts a second place feature value for recognizing a place related to the input image.
  • the recognition unit 25 determines a place related to the input image using the first place feature value and the second place feature value.
  • the recognition unit generates a synthesized place feature value by synthesizing the first place feature value and the second place feature value according to a predetermined criterion, and applies a predetermined connection weight to recognize the synthesized place feature value and place. To recognize the place.
  • FIG. 3 is a block diagram showing an artificial neural network-based learning apparatus according to an embodiment of the present invention.
  • the artificial neural network-based learning apparatus 3 includes a processor 31 and a memory 32 that perform signal processing for generating a correlation feature value based on the artificial neural network. And the processor 31 extracts object feature values for the training image, generates correlation feature values for objects and places using the object feature values and connection weights, and generates the connection weights through error backpropagation. Update.
  • the object feature values are elements of an n (here, n is a natural number) dimensional array
  • the object feature value vector calculation unit converts each of the elements into n normalized vector data.
  • the correlation feature value is a transformation parameter representing the object-place correlation, and a correlation feature value matrix is calculated. Thereafter, a multiplication of the correlation feature value matrix and the object feature value vector is performed to calculate a location feature value vector.
  • the processor calculates an error by comparing the generated location feature value with a previously stored location feature value for the training image, and performs signal processing for transferring the error to the correlation feature value generator.
  • a place prediction value is output using the calculated place feature value vector, and the loss value of the correlation feature value matrix is calculated by comparing the place prediction value and the place feature value.
  • the transformation parameter is adjusted using a back propagation algorithm based on the loss value.
  • the processor updates the transform model by adjusting the transform parameter until the loss value calculated by comparing the place prediction value and the place feature value is less than a threshold value.
  • the memory 32 may store programs (one or more instructions) for processing and controlling the processor 31, and programs stored in the memory 32 may be divided into a plurality of modules according to functions.
  • the memory 32 may store a location feature value according to an embodiment of the present invention.
  • the artificial neural network-based learning apparatus is for generating correlation feature values, and indicates an object-place correlation for place recognition related to a plurality of objects included in the analysis target image as a transformation parameter.
  • a transformation model based on an artificial neural network is generated, and the transformation model is updated by adjusting the transformation parameter based on the object feature value.
  • the apparatus for recognizing a place based on an artificial neural network generates a feature value matrix for correlation between an object and a place by using each classifier learned using the object and place recognition learning data of an image unit. can do.
  • a representative example of deriving correlations for two different types is the correlation coefficient of statistics.
  • this method when there is a correct answer about which object and place appear for all images, all correct answers are statistically analyzed and the relationship between a specific object and place is expressed in numerical terms.
  • information on the objects and places appearing in the video is required together.
  • existing datasets that provide a large amount of images, such as ImageNet and Places 2 only provide information on either object or place. Therefore, in order to apply to the existing dataset, information on an object or place that is not provided must be entered directly.
  • this method is difficult to use in the Convolutional Neural Network (CNN) structure, which is most often used for classifying objects or places in existing images.
  • CNN Convolutional Neural Network
  • the learning device 3 which can be implemented as an artificial neural network-based place recognition device 1 and a correlation feature value matrix generation algorithm generation device, according to an embodiment of the present invention, is provided in the form of an artificial neural network.
  • the correlation feature value matrix By using the correlation feature value matrix, there is an advantage that it can be applied to any artificial neural network structure for location recognition.
  • it can be modularized in the form of an artificial neural network capable of end-to-end learning to facilitate implementation with an artificial neural network.
  • Unlike the conventional statistical method which required both an object and a place in a single image, even if only one correct answer exists for an object or a place, it is possible to generate a correlation feature value matrix with a weak supervised learning algorithm of deep learning.
  • the processor when it extracts the object feature value for the training image, it receives the training image and extracts the object feature value for the training image from the training image using a convolutional neural network.
  • the correlation feature value represents a mutual relationship according to each combination of a plurality of predetermined objects and an object selected from a plurality of places and a place.
  • An artificial neural network-based learning apparatus generates a correlation feature value matrix between an object and a place by using image unit place recognition data, and modulates it into an artificial neural network structure to use a weak supervised learning algorithm. Create a matrix of values.
  • FIG. 4 is a diagram illustrating an exemplary structure including a convolutional neural network of a place recognition apparatus based on an artificial neural network according to an embodiment of the present invention.
  • the apparatus for recognizing a place based on an artificial neural network generates a matrix of correlation feature values between an object and a place by using image unit place recognition data and a CNN using a weak supervised learning method.
  • the present invention proposes a weak supervised learning method that can generate a matrix of correlation feature values between objects and places using the dataset. .
  • the input unit 210 receives an analysis target image 211 including the location 212 to be recognized. Thereafter, the image to be analyzed is input to the object feature value extraction unit.
  • the object feature value extraction unit 220 tracks a plurality of objects included in the analysis target image by using a convolutional neural network (CNN) previously learned in the object recognition dataset.
  • CNN convolutional neural network
  • the object recognition dataset is a set of regions in the analysis target image that are likely to detect objects related to the location to be recognized.
  • the artificial neural network implemented as a convolutional neural network includes a plurality of convolution modules 221, 222, 223, and the first convolution module includes first convolution layers 221a, 221b, 221c) and a first mapping unit (ReLU).
  • the first convolution layer may extract a first convolutional feature image by convolution filtering the first local image through a convolution operation.
  • the first mapping unit may be a rectified linear unit (ReLU) that is an activation function.
  • the first mapping unit (ReLU) maps the first convolutional feature image extracted through the first convolutional layer according to a predetermined function, so that the first convolutional feature image is linearized and activated. Can be calculated.
  • the deep learning layer (fully connected layer) 230 extracts information related to a plurality of tracked objects as object feature values.
  • the correlation feature value management unit 300 learns a correlation feature value representing a correlation according to a combination of an object and a place that can be selected from a plurality of predetermined objects and a plurality of places, or based on the already learned correlation feature value. Manage information about.
  • a separate learning device generates a transformation model based on an artificial neural network representing an object-place correlation for place recognition related to a plurality of objects included in the analysis target image as a transformation parameter, and the transformation based on the object feature values
  • the transformation model is updated by adjusting the parameters.
  • a place recognition dataset and a deep CNN trained on an object recognition dataset are required. It is preferable to use data sets that are publicly available as place-aware datasets, such as Places2, SUN 397, MIT indoor67, and Scene 15. In addition, it is desirable to use a structure that has been learned in advance for an object recognition dataset such as ImageNet, such as AlexNet, ResNet, and DenseNet, which are widely used in the past as Deep CNN.
  • ImageNet such as AlexNet, ResNet, and DenseNet
  • a place recognition data set for constructing an artificial neural network structure and data sets capable of implementing the function of an object recognition data set may be used.
  • the location information of the image is obtained using the image unit location recognition dataset, and the information of the object for the given image is extracted from Deep CNN. Using these two pieces of information, we train the correlation feature value matrix between objects and places.
  • X is the object feature value vector extracted from Deep CNN
  • the correlation feature value matrix between the object and place is M
  • the place feature value vector obtained through this correlation matrix is Y
  • the correlation feature value between the object is obtained using the matrix.
  • the dimensions of the variables used in [Equation 1] are respectively , , to be.
  • the object is first found in Deep CNN of the object information extraction unit, which is an object classifier.
  • this classifier cannot perform 100% perfect detection, it can be recognized that objects other than desks and blackboards exist, but their feature values will be very low. Therefore, when visualizing the recognition of a place with the extracted object feature values using this, it appears as shown in FIG. 4 below.
  • FIG. 5 is a diagram illustrating a process of calculating a location feature value vector of a learning device based on an artificial neural network according to an embodiment of the present invention.
  • FIG. 5 is a visualization of recognizing a place with the extracted object feature values, and when the calculated place feature value vector is added to the structure of an artificial neural network, learning is performed as shown in [Equation 2] below.
  • L is the loss value of the correlation feature matrix calculated while training the artificial neural network.
  • the correlation feature matrix M between objects and places is learned by the value of ⁇ M .
  • whether or not the ⁇ M value is calculated determines whether or not the artificial neural network can be trained.
  • the error can be backpropagated if the differential equation can be calculated for X , which was the input of the correlation feature value matrix between objects and places.
  • this differential equation can be calculated in [Equation 3]
  • the correlation feature value matrix indicates that there is no problem in backpropagating errors.
  • the correlation feature value matrix between the object and the place of the artificial neural network-based place recognition apparatus according to an embodiment of the present invention can be learned and errors can be backpropagated. Showed. Therefore, it can be seen that the proposed correlation feature value matrix can be modularized into the structure of an artificial neural network, and the entire artificial neural network structure using this can be learned end-to-end.
  • 6A and 6B are graphs showing a correlation weight of an object and a correlation weight of a place according to an analysis of a place recognition apparatus based on an artificial neural network according to an embodiment of the present invention.
  • the object and the place You can analyze how much impact it has.
  • FIG. 6A is a matrix of correlation feature values between objects and places The correlation weight of the object by adding the absolute value of Is calculated, and (b) of FIG. 6 is the correlation weight of the added place for all objects Is calculated, sorted in ascending order, and printed as a graph.
  • the data set used for Deep CNN training is ImageNet, and there are 1000 object types, and the data set used for the place to be classified is Places 2 and 365 place types.
  • objects affecting a place are evenly distributed without being biased to a specific object.
  • places it can be seen that certain places have a lot of influence on the object. Therefore, in order to check which places have a lot of influence on the object, 5 places with high weight can be extracted in descending order and output as shown in [Table 1] below.
  • a matrix of correlation feature values between objects and places can be created using only place recognition data. Unlike the previous statistical method, where information on the two types had to be present in one image to determine the correlation between the other two types, a dataset that provides information on one type is easily found as an open data set. I can. Therefore, if this method is used, it is possible to learn easily with the deep learning method without additional labor.
  • FIG. 7 is a diagram showing an exemplary structure including a convolutional neural network of a place recognition apparatus based on an artificial neural network according to another embodiment of the present invention.
  • an exemplary structure including a convolutional neural network of a place recognition device based on an artificial neural network is constructed, and is not limited to a specific CNN, but is widely used in the past to show that it has been generalized. It is preferable to apply to AlexNet, ResNet-18, ResNet-50, and DenseNet-161. In addition, the results are shown in [Table 2].
  • FIG. 7 As shown in [Table 2], it can be seen that the performance increases in all CNNs.
  • the structure of FIG. 7 is very easy for learning an artificial neural network because end-to-end learning is possible. If the method proposed in FIG. 7 is changed to a correlation coefficient of statistics, end-to-end learning is not possible, and an additional artificial neural network module suitable for the correlation coefficient must be added. That is, the proposed method has the advantage that it can be used without an additional module and can be applied to any CNN.
  • a computer-readable recording medium in which a program for executing an artificial neural network-based location recognition device recognition method on a computer is recorded.
  • Such a computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination.
  • the program instructions recorded on the recording medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in computer software.
  • Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floppy disks.
  • -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like.
  • Examples of the program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.
  • the above-described hardware device may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

Landscapes

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

Abstract

La présente invention concerne un dispositif de reconnaissance de lieu basé sur un réseau neuronal artificiel et son dispositif d'apprentissage qui comprennent une unité de gestion de valeur de caractéristique de corrélation basée sur un réseau neuronal artificiel, apprenant une valeur de caractéristique de corrélation indiquant une corrélation conformément à une combinaison d'un objet et d'un lieu qui pourraient être sélectionnés parmi une pluralité d'objets prédéfinis et une pluralité de lieux prédéfinis ou gérant des informations relatives à une valeur de caractéristique de corrélation qui a déjà été apprise et améliorent les performances de reconnaissance de lieu d'unité d'image.
PCT/KR2020/001018 2019-04-09 2020-01-21 Dispositif de reconnaissance de lieu basé sur un réseau neuronal artificiel et son dispositif d'apprentissage Ceased WO2020209487A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
KR10-2019-0041544 2019-04-09
KR1020190041544A KR102211842B1 (ko) 2019-04-09 2019-04-09 상관관계 점수 행렬 생성 알고리즘을 이용한 인공 신경망 기반의 장소 인식 장치 및 방법

Publications (1)

Publication Number Publication Date
WO2020209487A1 true WO2020209487A1 (fr) 2020-10-15

Family

ID=72751920

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/KR2020/001018 Ceased WO2020209487A1 (fr) 2019-04-09 2020-01-21 Dispositif de reconnaissance de lieu basé sur un réseau neuronal artificiel et son dispositif d'apprentissage

Country Status (2)

Country Link
KR (1) KR102211842B1 (fr)
WO (1) WO2020209487A1 (fr)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR102672247B1 (ko) * 2021-06-30 2024-06-05 주식회사 카카오 광고의 예상 과금액을 예측하는 방법 및 장치
KR102681856B1 (ko) * 2021-11-22 2024-07-03 광운대학교 산학협력단 상황 정보 기반의 객체 검출 정확도 보정 방법 및 시스템

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101551237B1 (ko) * 2014-01-07 2015-09-09 한양대학교 산학협력단 직선 및 기하관계 기반 실외 장소 인식 방법 및 시스템
KR20170030332A (ko) * 2015-09-09 2017-03-17 한국과학기술연구원 객체 고유의 구조적 특징을 이용한 객체 인식 장치 및 그 방법

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101551237B1 (ko) * 2014-01-07 2015-09-09 한양대학교 산학협력단 직선 및 기하관계 기반 실외 장소 인식 방법 및 시스템
KR20170030332A (ko) * 2015-09-09 2017-03-17 한국과학기술연구원 객체 고유의 구조적 특징을 이용한 객체 인식 장치 및 그 방법

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
KIM, SUNGHO: "Collaborative Place and Object Recognition in Video Using Bidirection Context Information", THE JOURNAL OF KOREA ROBOTICS SOCIETY, vol. 1, no. 2, December 2006 (2006-12-01), pages 432 - 439, XP055748641 *
OH, SUJIN: "Real-Time Place Recognition System Design and Implementation", MASTER'S THESIS, GRADUATE SCHOOL OF EWHA WOMANS UNIVERSITY, 2007, pages 6 *
ZETAO CHEN: "Deep Learning Features at Scale for Visual Place Recognition", 2017 IEEE INTERNATIONALCONFERENCE ON ROBOTICS AND AUTOMATION (ICRA, 29 May 2017 (2017-05-29), Singapore, pages 3223 - 3230, XP033127155, DOI: 10.1109/ICRA.2017.7989366 *

Also Published As

Publication number Publication date
KR20200120987A (ko) 2020-10-23
KR102211842B1 (ko) 2021-02-02

Similar Documents

Publication Publication Date Title
Torralba et al. Contextual models for object detection using boosted random fields
CN101625675B (zh) 信息处理装置、信息处理方法和计算机程序
WO2021230457A1 (fr) Procédé d'apprentissage et dispositif d'apprentissage pour entraîner un réseau de détection d'objets à l'aide de cartes d'attention et procédé de test et dispositif de test le mettant en œuvre
WO2022065817A1 (fr) Procédés d'entraînement et de test d'un réseau d'obscurcissement susceptible d'exécuter des processus de dissimulation distincts pour des régions distinctes d'une image originale, et dispositifs d'entraînement et de test les utilisant
CN116935447B (zh) 基于自适应师生结构的无监督域行人重识别方法及系统
CN112465144B (zh) 基于有限知识的多模态示范意图生成方法及装置
WO2017039086A1 (fr) Système de modularisation d'apprentissage profond sur la base d'un module d'extension internet et procédé de reconnaissance d'image l'utilisant
WO2022163996A1 (fr) Dispositif pour prédire une interaction médicament-cible à l'aide d'un modèle de réseau neuronal profond à base d'auto-attention, et son procédé
CN114782752B (zh) 基于自训练的小样本图像集成分类方法及装置
WO2017164478A1 (fr) Procédé et appareil de reconnaissance de micro-expressions au moyen d'une analyse d'apprentissage profond d'une dynamique micro-faciale
WO2023096254A1 (fr) Système de mise en correspondance d'emploi sur la base de l'intelligence artificielle
WO2023027340A1 (fr) Procédé de formation et de test d'un réseau de dissimulation susceptible de dissimuler des données pour protéger des informations personnelles, et dispositif d'apprentissage et dispositif d'essai l'utilisant
CN110348505B (zh) 车辆颜色分类模型训练方法、装置和车辆颜色识别方法
WO2018212584A2 (fr) Procédé et appareil de classification de catégorie à laquelle une phrase appartient à l'aide d'un réseau neuronal profond
WO2021246810A1 (fr) Procédé d'entraînement de réseau neuronal par auto-codeur et apprentissage multi-instance, et système informatique pour la mise en oeuvre de ce procédé
WO2020246655A1 (fr) Procédé de reconnaissance de situation et dispositif permettant de le mettre en œuvre
CN118569305A (zh) 一种基于超图神经网络的多模态轨迹预测方法及装置
WO2020209487A1 (fr) Dispositif de reconnaissance de lieu basé sur un réseau neuronal artificiel et son dispositif d'apprentissage
WO2024029669A1 (fr) Système de reconstruction d'image antagoniste et procédé de reconstruction d'image antagoniste
WO2019035544A1 (fr) Appareil et procédé de reconnaissance faciale par apprentissage
CN115187898A (zh) 一种面向大场景多对象的自监督群体检测方法及装置
CN111191461A (zh) 一种基于课程学习的远程监督关系抽取方法
CN113822161A (zh) 肤色与三级背景更新模型融合的动态手部检测方法及装置
CN119863823A (zh) 一种基于超图驱动软语义动态学习的可见光-红外行人重识别方法
WO2023249184A1 (fr) Système d'apprentissage adverse et procédé d'apprentissage adverse

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: 20787328

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20787328

Country of ref document: EP

Kind code of ref document: A1