WO2019127451A1

WO2019127451A1 - Image recognition method and cloud system

Info

Publication number: WO2019127451A1
Application number: PCT/CN2017/120087
Authority: WO
Inventors: 李少华; 刘兆祥; 廉士国
Original assignee: 深圳前海达闼云端智能科技有限公司
Priority date: 2017-12-29
Filing date: 2017-12-29
Publication date: 2019-07-04
Also published as: CN108235770B; CN108235770A

Abstract

Provided are an image recognition method and a cloud system. The method comprises: using a preset feature extraction network to acquire a feature vector of an image to be recognized; and using a preset classifier to obtain, according to a Euclidean distance between the feature vector of the image to be recognized and a class center point corresponding thereto, a recognition result of the image to be recognized. The present application is based on an improved classifier, and increases the aggregation degree of feature vectors within a class, expands the distinguishability of feature vectors between classes, and also improves the robustness of the cloud system.

Description

Image recognition method and cloud system

Technical field

The present application relates to the field of image recognition technologies, and in particular, to an image recognition method and a cloud system.

Background technique

Artificial Neural Network (ANN) is suitable for application scenarios such as feature classification. Usually, linear classifiers are used to classify features. The most widely used linear classifier is the Softmax classifier. For a classification task with category n, the feature vector x _i (i=1,...,m) extracted by the ANN belongs to the jth (j=1,...l, The probability P _ij of the class n) is:

During the training process, the loss function is defined as follows:

Where w and b are neural network parameters, w is the connection weight, b is the offset, 1{·} is the explicit function, when the expression is true, 1{expression}=1, when the expression When it is false, 1{expression}=0.

The classification result of the Softmax classifier depends on the feature vector and the inner product representing the class center vector. When the feature vector extracted by the ANN is optimized by the loss function defined by the Softmax classifier and the cross entropy, only the feature vector can be linearly separable. However, only linearly separable feature vectors have the following problems in practical applications:

1) For eigenvectors close to the category boundary, it is susceptible to misclassification caused by small disturbances, and the system is less robust;

2) For non-classification tasks (for example, face recognition), the extracted feature vectors cannot guarantee good intra-class aggregation and inter-class discrimination.

Summary of the invention

The image recognition method and the cloud system are proposed in the embodiment of the present application to solve the problem that the existing classifier has low recognition degree of the feature vector close to the category boundary, and the extracted feature vector cannot guarantee good intra-class aggregation degree and inter-class distinction. Technical problems.

In one aspect, an embodiment of the present application provides an image recognition method, including:

Acquiring a feature vector of the image to be identified by using a preset feature extraction network;

The recognition result of the image to be recognized is obtained according to the Euclidean distance of the feature vector of the image to be recognized and the corresponding class center point by using a preset classifier.

In another aspect, an embodiment of the present application provides an image recognition cloud system, including:

a feature extraction network, configured to acquire a feature vector of the image to be identified by using a preset feature extraction network;

The classifier is configured to obtain, by using a preset classifier, a recognition result of the image to be identified according to a feature distance of the feature vector of the image to be recognized and a corresponding Euclidean distance of the class center point.

In another aspect, an embodiment of the present application provides an electronic device, where the electronic device includes:

Transceiver, memory, one or more processors;

One or more modules, the one or more modules being stored in the memory and configured to be executed by the one or more processors, the one or more modules comprising Instructions for each step.

In another aspect, embodiments of the present application provide a computer program product for use with an electronic device, the computer program product comprising a computer readable storage medium and a computer program mechanism embedded therein, the computer program mechanism Instructions are included for performing the various steps in the above methods.

The benefits are as follows:

In this embodiment, the feature vector of the image to be recognized is acquired by using the trained feature extraction network, and the image to be recognized is obtained according to the Euclidean distance of the feature vector of the image to be recognized and the corresponding class center point by using the trained classifier. Identification result. By improving the existing classifier, the degree of aggregation of feature vectors in the class is increased, the distinguishability of feature vectors between classes is expanded, and the robustness of the cloud system is improved.

DRAWINGS

Specific embodiments of the present application will be described below with reference to the accompanying drawings, in which:

1 is a schematic diagram of a method for image recognition in Embodiment 1 of the present application;

2 is a schematic diagram of classification of a classifier in an image recognition method according to Embodiment 1 of the present application;

3 is a structural diagram of a cloud system for image recognition in Embodiment 2 of the present application;

FIG. 4 is a schematic structural diagram of an electronic device according to Embodiment 3 of the present application.

Detailed ways

The essence of the technical solution of the embodiment of the present invention is further clarified by specific examples below.

The exemplary embodiments of the present application are further described in detail below with reference to the accompanying drawings, in which the embodiments described are only a part of the embodiments of the present application, but not all embodiments. An exhaustive example. And in the case of no conflict, the features in the embodiments and the embodiments in the description can be combined with each other.

The inventor noticed during the invention:

The existing Softmax classifier measures the similarity based on the inner product of the feature vector, and the intuitiveness is poor. For eigenvectors that are only linearly separable and close to the class boundary, the existing Softmax classifiers are highly susceptible to misclassification caused by small disturbances, and the robustness is low. For non-classification tasks, the extracted feature vectors cannot guarantee good classes. Internal aggregation and inter-class discrimination.

In view of the above-mentioned deficiencies/based on the above, the embodiment of the present application proposes to improve the manner in which the classifier measures the similarity by means of the inner product calculation to the method of calculating the Euclidean distance, that is, based on the existing Softmax classifier. The loss function of the class center vector is improved to the loss function based on the class center point, so as to increase the feature vector aggregation degree in the class and expand the technical effect of the distinguishability of the feature vector between classes.

In order to facilitate the implementation of the present application, the following examples are described.

Example 1

FIG. 1 is a schematic diagram of a method for image recognition in the first embodiment of the present application. As shown in FIG. 1, the method includes:

Step 101: Acquire a feature vector of the image to be identified by using a preset feature extraction network.

Step 102: Using a preset classifier, obtaining a recognition result of the image to be recognized according to the Euclidean distance of the feature vector of the to-be-identified image and its corresponding class center point.

In the implementation, the execution body of the above step may be a cloud server, and the training device in the cloud server trains the ANN based on the image sample, the image tag, and the modified loss function of the modified Softmax classifier, and in the process of training optimization, The loss function deviates the parameters of each layer of the ANN, and uses the backward conduction algorithm to optimize the parameters of each layer of the ANN, so that the trained ANN can obtain the feature vector of the image to be identified by extracting the network, and use the modified Softmax classification. The feature vector of the image to be recognized is recognized, and the recognition result of the image to be recognized is obtained.

In this embodiment, the initialized feature extraction network and the classifier are trained by using the first loss function to obtain a preset feature extraction network and a classifier, and the first loss function L is:

Said

Where L _i is the loss function of the feature vector x _i (i=1, . . . , m) of the image sample, and C _j is the jth (j=1, . . . , n) corresponding to the feature vector x _i of the image sample. The center point of the image category, M is the preset first neural network parameter.

In this embodiment, the preset condition of the neural network parameter M is:

M times the Euclidean distance of feature vectors of sample images x _i corresponding to the image category of the j-th class is less than or equal to the center point of the sample image feature vector x _i European center point and one of the other classes of image categories Distance; or,

The probability that the feature vector x _{i of} the image sample belongs to the j-th image category is greater than or equal to the sum of the probability that the feature vector x _{i of} the image sample belongs to any other image category and the preset second neural network parameter δ.

In the implementation, the setting of the loss function of the initialized modified Softmax classifier specifically includes:

The existing Softmax classifier is improved based on the class center vector metric similarity to the similarity based on the class center point metric, that is, the metric of the similarity based on the Euclidean distance of the feature vector x _i and its corresponding class center point, the eigenvector The probability that x _i (i=1,...,m) belongs to the jth class is:

During the training process, the loss function is defined as follows:

Where C _j is the central point of the jth (j=1,...l,n) class. The loss function at this time cannot effectively improve the aggregation degree of the feature vector, and introduces a new parameter M for the above loss function, and the improved loss. The function is:

Wherein, the value of M should satisfy the condition that M times of the Euclidean distance of the feature vector x _i and its corresponding class center point is less than or equal to the Euclidean distance of the feature vector x _i and any other type of center point, or the feature vector x _i The probability of belonging to the jth class is greater than or equal to the sum of the probability that the feature vector x _i belongs to the class 1 and the neural network parameter δ, that is,

or

P _ij ≥P _il +δ δ>0, l≠i.

2 is a schematic diagram of classification of a classifier in an image recognition method according to Embodiment 1 of the present application. As shown in FIG. 2, a new parameter M is introduced in a modified Softmax classifier, so that M=2, visible, introduced in the loss function. The new parameter M can further increase the degree of aggregation of feature vectors within the class and expand the distinguishability of feature vectors between classes.

In this embodiment, the method further includes:

The initial feature extraction network and the classifier are trained by using the first loss function to obtain a first feature extraction network and a preset classifier;

The first feature extraction network is trained by using a preset second loss function to obtain a preset feature extraction network.

In this embodiment, the preset second loss function L _C is:

Where C _j is the center point of the j-th image category corresponding to the feature vector x _i (i=1, . . . , m) of the image sample, and the class center point of the second loss function is related to the first loss function The class center points are the same.

In the implementation, if the new parameter M introduced in the loss function of the modified Softmax classifier is 1, the modified Softmax classifier loss function does not take into account the "safe range" of the image category boundary, and can be step-by-step trained. The method realizes the optimization of the parameters of each layer of the ANN, thereby achieving the technical effect of increasing the degree of feature vector aggregation within the class and expanding the distinguishability of the feature vector between the classes. The training process specifically includes:

1) Perform the first stage training on the initialized ANN feature extraction network and the modified Softmax classifier. According to the feature vector of the image sample extracted by the ANN feature extraction network and the preset image tag, the loss value is calculated by using the forward conduction algorithm and the modified loss function of the Softmax classifier, and the first loss function is used for the parameters of the ANN layer. The partial derivative is used to perform the first stage training on the initialized ANN feature extraction network and the modified Softmax classifier. The trained first ANN feature extraction network and the modified Softmax are obtained by optimizing the parameters of the various layers of the ANN. Classifier.

2) Perform the second stage training on the first ANN feature extraction network. The fixed-corrected Softmax classifier, that is, keeps the class center C of each category in the ANN classification task unchanged, and uses the set second loss function to train the first ANN feature extraction network to obtain a trained ANN feature extraction network. Specifically, setting the second loss function L _C is:

The second loss function L _{C is used} to derive the eigenvector x _i of the image sample and the ANN feature extraction network layer parameters, and the backward conduction algorithm is used to optimize the parameters of the ANN feature extraction network layer to make the trained ANN features The extraction network has higher accuracy when extracting the feature vector of the image to be recognized, that is, the degree of aggregation of the extracted similar feature vectors is higher.

The present application provides a detailed description of Embodiment 1 of the present application by taking a specific scenario as an example.

The application scope of the embodiments of the present application includes, but is not limited to, ANN-based face image recognition, and the face image recognition based on ANN is taken as an example. The specific process is as follows:

The training process of the ANN feature extraction network and the modified Softmax classifier:

Step 201: Perform the first stage training by the initialized ANN feature extraction network and the modified Softmax classifier. According to the feature vector of the image sample extracted by the ANN feature extraction network and the preset image tag, the loss value is calculated by using the forward conduction algorithm and the modified loss function of the Softmax classifier, and the first loss function is used for the parameters of the ANN layer. The partial derivative is obtained by using the backward conduction algorithm to optimize the parameters of each layer of the ANN, and the first feature extraction network and the modified Softmax classifier are obtained. The first loss function is defined as follows:

Where L _i is the loss function of the feature vector x _i (i=1, . . . , m) of the image sample, and C _j is the jth (j=1, . . . , n) corresponding to the feature vector x _i of the image sample. The class center point of the image category, M is the preset first neural network parameter.

Step 202: If the new parameter M introduced in the loss function of the modified Softmax classifier is 1, the second stage training is performed on the first ANN feature extraction network. Specifically include:

Keep the class center C of each category in the ANN classification task unchanged, and use the set second loss function to train the first ANN feature extraction network to obtain the trained ANN feature extraction network, that is, do not do the modified Softmax classifier. Training, only the second feature optimization of the ANN feature extraction network layer parameters, the second loss function is defined as follows:

The second loss function L _{C is used} to derive the eigenvector x _i of the image sample and the ANN feature extraction network layer parameters, and the backward conduction algorithm is used to optimize the parameters of the ANN feature extraction network layer to obtain the trained ANN feature extraction. The internet.

Based on the trained ANN feature extraction network and the modified Softmax classifier recognition process:

Step 203: Acquire an image to be identified, extract a feature vector of the image to be recognized by using the trained ANN feature extraction network, and identify a feature vector of the image to be recognized by using the modified Softmax classifier to obtain a recognition result of the image to be identified.

Example 2

Based on the same inventive concept, an image recognition cloud system is also provided in the embodiment of the present application. Since the principle of solving the problem of these devices is similar to an image recognition method, the implementation of these devices can refer to the implementation of the method, and the repetition is not Let me repeat.

FIG. 3 is a schematic diagram of a cloud system architecture for image recognition in the second embodiment of the present application. As shown in FIG. 3, the image recognition cloud system 300 may include:

a feature extraction network 301, configured to acquire a feature vector of an image to be identified by using a preset feature extraction network;

The classifier 302 is configured to obtain a recognition result of the image to be recognized according to the Euclidean distance of the feature vector of the to-be-identified image and the corresponding class center point by using a preset classifier.

The training device 303 is configured to train the initialized feature extraction network and the classifier by using the first loss function to obtain a preset feature extraction network and a classifier, where the first loss function L is:

Said

In this embodiment, the preset condition of the neural network parameter M is:

In this embodiment, the training device 303 is further configured to use the first loss function to train the initialized feature extraction network and the classifier to obtain a first feature extraction network and a preset classifier;

In this embodiment, the preset second loss function L _C is:

Where C _j is a class center point of the jth image class corresponding to the feature vector x _i (i=1, . . . , m) of the image sample, and the class center point of the second loss function and the first loss function The class center points are the same.

Example 3

Based on the same inventive concept, an electronic device is also provided in the embodiment of the present application. Since the principle is similar to an image recognition method, the implementation of the method may refer to the implementation of the method, and the repeated description is not repeated.

4 is a schematic structural diagram of an electronic device in Embodiment 3 of the present application. As shown in FIG. 4, the electronic device includes: a transceiver device 401, a memory 402, one or more processors 403, and one or more modules. The one or more modules are stored in the memory and configured to be executed by the one or more processors, the one or more modules including steps for performing the steps of any of the above methods instruction.

Example 4

Based on the same inventive concept, the embodiment of the present application further provides a computer program product used in combination with an electronic device. Since the principle is similar to an image recognition method, the implementation of the method can refer to the implementation of the method, and the details are not repeated. . The computer program product comprises a computer readable storage medium and a computer program mechanism embodied therein, the computer program mechanism comprising instructions for performing the various steps of any of the above methods.

For the convenience of description, the various parts of the above-described apparatus are separately described by functions into various modules. Of course, the functions of each module or unit may be implemented in the same software or hardware in the implementation of the present application.

Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system, or computer program product. Thus, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment in combination of software and hardware. Moreover, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (system), and computer program products according to embodiments of the present application. It will be understood that each flow and/or block of the flowchart illustrations and/or FIG. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing device to produce a machine for the execution of instructions for execution by a processor of a computer or other programmable data processing device. Means for implementing the functions specified in one or more of the flow or in a block or blocks of the flow chart.

The computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device. The apparatus implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.

These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on a computer or other programmable device to produce computer-implemented processing for execution on a computer or other programmable device. The instructions provide steps for implementing the functions specified in one or more of the flow or in a block or blocks of a flow diagram.

While the preferred embodiment of the present application has been described, it will be apparent that those skilled in the art can make further changes and modifications to the embodiments. Therefore, it is intended that the appended claims be interpreted as

Claims

An image recognition method, comprising:

Acquiring a feature vector of the image to be identified by using a preset feature extraction network;

The recognition result of the image to be recognized is obtained according to the Euclidean distance of the feature vector of the image to be recognized and the corresponding class center point by using a preset classifier.
The method according to claim 1, further comprising: training the initialized feature extraction network and the classifier with a first loss function to obtain a preset feature extraction network and a classifier, the first loss function L is:

Said

Where L i is the loss function of the feature vector x i (i=1, . . . , m) of the image sample, and C j is the jth (j=1, . . . , n) corresponding to the feature vector x i of the image sample. The center point of the image category, M is the preset first neural network parameter.
The method according to claim 2, wherein the predetermined condition of the neural network parameter M is:

M times the Euclidean distance of feature vectors of sample images x i corresponding to the image category of the j-th class is less than or equal to the center point of the sample image feature vector x i European center point and one of the other classes of image categories Distance; or,

The probability that the feature vector x i of the image sample belongs to the j-th image category is greater than or equal to the sum of the probability that the feature vector x i of the image sample belongs to any other image category and the preset second neural network parameter δ.
The method of claim 1 or 2, further comprising:

The initial feature extraction network and the classifier are trained by using the first loss function to obtain a first feature extraction network and a preset classifier;

The first feature extraction network is trained by using a preset second loss function to obtain a preset feature extraction network.
The method of claim 4 wherein said predetermined second loss function L C is:

Where C j is a class center point of the jth image class corresponding to the feature vector x i (i=1, . . . , m) of the image sample, and the class center point of the second loss function and the first loss function The class center points are the same.
An image recognition cloud system, comprising:

a feature extraction network, configured to acquire a feature vector of the image to be identified by using a preset feature extraction network;

The classifier is configured to obtain, by using a preset classifier, a recognition result of the image to be identified according to a feature distance of the feature vector of the image to be recognized and a corresponding Euclidean distance of the class center point.
The cloud system according to claim 6, further comprising a trainer for training the initialized feature extraction network and the classifier by using the first loss function to obtain a preset feature extraction network and classification The first loss function L is:

Said

Where L i is the loss function of the feature vector x i (i=1, . . . , m) of the image sample, and C j is the jth (j=1, . . . , n) corresponding to the feature vector x i of the image sample. The center point of the image category, M is the preset first neural network parameter.
The cloud system according to claim 7, wherein the preset condition of the neural network parameter M is:

M times the Euclidean distance of feature vectors of sample images x i corresponding to the image category of the j-th class is less than or equal to the center point of the sample image feature vector x i European center point and one of the other classes of image categories Distance; or,

The probability that the feature vector x i of the image sample belongs to the j-th image category is greater than or equal to the sum of the probability that the feature vector x i of the image sample belongs to any other image category and the preset second neural network parameter δ.
The cloud system according to claim 6 or 7, further comprising a trainer for training the initialized feature extraction network and the classifier by using the first loss function to obtain the first feature extraction network and Preset classifier; and,

The first feature extraction network is trained by using a preset second loss function to obtain a preset feature extraction network.
The cloud system according to claim 9, wherein the preset second loss function L C is:

Where C j is a class center point of the jth image class corresponding to the feature vector x i (i=1, . . . , m) of the image sample, and the class center point of the second loss function and the first loss function The class center points are the same.
An electronic device, comprising:

Transceiver, memory, one or more processors;

One or more modules stored in the memory and configured to be executed by the one or more processors, the one or more modules including for performing claim 1 The instructions of the various steps in any of the methods described in 5.
A computer program product for use in conjunction with an electronic device, the computer program product comprising a computer readable storage medium and a computer program mechanism embedded therein, the computer program mechanism comprising means for performing any of claims 1-5 An instruction for each step in the method.