Pedestrian re-identification method and device and storage medium

By introducing a target backbone network that integrates a normalization layer and a simulated normalization layer into the pedestrian re-identification method, the problem of low recognition efficiency between different source domain datasets is solved, and high accuracy and high efficiency of pedestrian re-identification are achieved in different source domains.

CN118968548APending Publication Date: 2024-11-15UESTC (SHENZHEN) ADVANCED RES INST
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410994483.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing pedestrian re-identification methods have low generalization performance across datasets from different source domains, resulting in low identification efficiency. Furthermore, supervised methods are time-consuming and labor-intensive to train and involve personal privacy issues.

Method used

A target backbone network with a fusion normalization layer and a simulated normalization layer is adopted. The model is trained through two forward propagations and one backpropagation. The statistical parameters of instance normalization and batch normalization are combined to improve the model's generalization ability in different source domains.

Benefits of technology

This study improved the accuracy and efficiency of person re-identification on image datasets from different source domains, enhanced the model's generalization ability, and improved the accuracy and efficiency of identification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968548A_ABST
    Figure CN118968548A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of image recognition, in particular to a pedestrian re-recognition method, and the method comprises the steps: obtaining a to-be-recognized image and a target recognition model which comprises a first target backbone network with L fusion normalization layers and a second target backbone network with L simulation normalization layers; performing fusion normalized feature extraction on the to-be-recognized image through the first target backbone network to obtain a first target feature vector of the to-be-recognized image, and performing simulation normalized feature extraction on the to-be-recognized image through the second target backbone network to obtain a second target feature vector of the to-be-recognized image; obtaining a recognition feature vector of the to-be-recognized image according to the first target feature vector and the second target feature vector; and performing similarity measurement processing on the identification feature vector and a target feature vector in the target domain to obtain an identification result of a target pedestrian corresponding to the to-be-identified image. According to the method, the generalization ability of the pedestrian re-identification technology is improved, and the pedestrian re-identification efficiency is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a pedestrian re-identification method, device and storage medium. Background Art

[0002] Person re-identification is an important and challenging research problem in computer vision. It associates a specific pedestrian captured by a camera group at different times and physical locations, thereby quickly and accurately obtaining pedestrian-related information from different perspectives. Pedestrian re-identification technology lays the foundation for its upstream application scenarios, such as pedestrian tracking, pedestrian attribute recognition, and pedestrian behavior analysis. At present, pedestrian re-identification technology has shown great development prospects and economic benefits in many fields such as intelligent video surveillance, smart city transportation, smart photo albums, and smart retail.

[0003] Existing person re-identification methods are all trained and tested on the same dataset (i.e., collected from the same camera network), which means that the data sources follow the same distribution. However, in actual scenarios, there are large differences (i.e., domain differences) between different person re-identification datasets. For example, factors such as the scene perspective of the shooting, the light intensity of the shooting point, the resolution of the camera, and the clothing style of the pedestrians result in these supervised methods being only applicable to test sets with similar distributions to the current training set, and cannot be generalized to other datasets. Moreover, if re-training is required for the current scenario, the acquisition of identity label information required for training such supervised methods is time-consuming and laborious and involves personal privacy issues, which greatly reduces the practical application value of such methods. Therefore, the existing person re-identification methods have low performance in generalizing to datasets of different source domains, which leads to the problem of low efficiency of person re-identification. Summary of the invention

[0004] The embodiments of the present application provide a pedestrian re-identification method, device and storage medium to solve the technical problem in the prior art that the performance of data sets generalized to different source domains is low, resulting in low pedestrian re-identification efficiency. The embodiments of the present application achieve technical effects such as being compatible with image data sets from different source domains, improving the generalization ability of pedestrian re-identification technology, enhancing the accuracy of pedestrian re-identification, and thereby improving the efficiency of pedestrian re-identification.

[0005] In a first aspect, an embodiment of the present invention provides a pedestrian re-identification method, comprising:

[0006] Acquire an image to be recognized and a target recognition model, wherein the target recognition model includes a first target backbone network having L layers of fusion normalization layers and a second target backbone network having L layers of simulation normalization layers, wherein the L layers of fusion normalization layers correspond one to one to the L layers of simulation normalization layers, and L≥1;

[0007] Firstly, the first target backbone network is used to perform fusion normalized feature extraction on the image to be identified to obtain a first target feature vector of the image to be identified, and then the second target backbone network is used to perform simulated normalized feature extraction on the image to be identified to obtain a second target feature vector of the image to be identified;

[0008] Obtaining an identification feature vector of the image to be identified according to the first target feature vector and the second target feature vector;

[0009] The recognition feature vector is processed with a target feature vector in a target domain to measure similarity, so as to obtain a recognition result of a target pedestrian corresponding to the image to be recognized, wherein the target domain is a data set containing multiple target pedestrians.

[0010] Preferably, before obtaining the target recognition model, the method includes:

[0011] Obtain K source domains and an initial recognition model, wherein the initial recognition model includes a first backbone network having the L-layer fusion normalization layer and a second backbone network having the L-layer simulation normalization layer, one source domain is a training image set of a target pedestrian collected from the same camera network, and K≥2;

[0012] Iteratively training the initial recognition model through the K source domains until the number of iterative trainings meets an iteration number threshold, and determining the recognition model that meets the iteration number threshold as the target recognition model;

[0013] wherein, in each iterative training of the initial recognition model, the affine parameters of the normalization layer of the initial recognition model are controlled to be in a non-updated state, the first backbone network is used to perform a first forward propagation process on the input data corresponding to each iterative training to obtain a first prediction result of the input data, and first loss data of the first forward propagation is obtained based on the first prediction result, and the input data is a preset number of training images in a source domain of the K source domains under each iterative training;

[0014] Performing a second forward propagation process on the input data through the second backbone network to obtain a second prediction result of the input data, and obtaining second loss data of the second forward propagation based on the second prediction result;

[0015] According to the first loss data and the second loss data, the total loss data of the input data is obtained, and based on the total loss data, the learnable parameters in the initial recognition model are updated through back propagation processing.

[0016] Preferably, performing a first forward propagation process on the input data corresponding to each iterative training through the first backbone network to obtain a first prediction result of the input data includes:

[0017] Performing fusion normalization feature extraction on the input data through the first backbone network to obtain a first feature vector, and then performing classification processing on the first feature vector through a first classifier to obtain the first prediction result;

[0018] The process of extracting the fused normalized features in each fused normalized layer is as follows:

[0019]

[0020] i is the ordinal variable of the source domain, represents the input data of the fused normalized feature extraction, express The output data after fusion normalization feature extraction, that is, the first feature vector of each fusion normalization layer, i∈(1,2,...,K), K is the total number of source domains, γ RN represents the non-learnable scaling parameter in the fused normalized feature extraction, β RN represents the non-learnable translation parameter in the fusion normalization feature extraction, ∈ is a constant to prevent the denominator from being 0, Corresponding to input The mean and variance of the statistical parameters after instance normalization, Corresponding to input Mean and variance of statistical parameters after batch normalization.

[0021] Preferably, after the first backbone network performs fusion normalization feature extraction on the input data, the method further includes:

[0022] In each fusion normalization layer, based on the mean after the batch normalization processing and the variance after the batch normalization processing, the mean and variance of the accumulated parameters in the buffer module corresponding to the input data are updated, as shown in the following expression:

[0023]

[0024] in, is the mean of the cumulative parameter, is the variance of the accumulated parameters, and λ is the update weight set.

[0025] Preferably, performing a second forward propagation process on the input data through the second backbone network to obtain a second prediction result of the input data includes:

[0026] Performing simulated normalized feature extraction on the input data set through the second backbone network to obtain a second feature vector, and then performing classification processing on the second feature vector through a second classifier to obtain the second prediction result;

[0027] The process of extracting the simulated normalized features in each simulated normalized layer is as follows:

[0028]

[0029] j is the ordinal variable of the source domain, j∈(1,2,...,K) and i≠j, represents the input data for simulated normalized feature extraction, Represents input The output data after simulated normalization feature extraction, i.e., the second feature vector of each simulated normalization layer, γ EN represents the non-learnable scaling parameter in simulated normalization feature extraction, β EN represents the non-learnable translation parameter in simulated normalization feature extraction, yes The mean and variance of the statistical parameters after instance normalization, It comes from the jth buffer module Buffer j The mean and variance of the cumulative parameters stored in a j , b j Respectively represent Buffer j The two learnable parameters are used to adaptively learn the statistical parameters and cumulative parameters in the corresponding instance normalization processing.

[0030] Preferably, the first loss data includes: the ternary loss of the first backbone network and the classification loss of the first classifier The second loss data includes: the ternary loss of the second backbone network and the classification loss of the second classifier The total loss data

[0031] Preferably, the affine parameters include: a scaling parameter γ and a translation parameter β of a normalization layer in the initial recognition model.

[0032] Preferably, the target recognition model and the initial recognition model are both ReNorm models, the first backbone network is a backbone network in which all batch normalization processing in the ResNet50 network is replaced by a fused normalization layer formed by fused normalization feature extraction, and the second backbone network is a backbone network in which all batch normalization processing in the ResNet50 network is replaced by a simulated normalization layer formed by simulated normalization feature extraction.

[0033] Based on the same inventive concept, in a second aspect, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of a pedestrian re-identification method when executing the program.

[0034] Based on the same inventive concept, in a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the pedestrian re-identification method are implemented.

[0035] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages:

[0036] Based on the pedestrian re-identification method of the embodiment of the present invention, the first target backbone network with a fusion normalization layer RN (RemixNormalization) performs the first forward propagation processing on the image to be identified, and extracts the first target feature vector of the image to be identified. Then, the second target backbone network with an emulation normalization layer EN (Emulation Normalization) performs the second forward propagation processing on the image to be identified, and extracts the second target feature vector of the image to be identified. Among them, the fusion normalization layer RN (Remix Normalization) combines the existing instance normalization IN (Instance Normalization) and batch normalization BN (BatchNormalization) to normalize the statistical parameters. Batch normalization BN can extract discriminative features from the image. Instance normalization IN can eliminate the characteristics of domain information to a certain extent. The simulated normalization layer EN adopts a structure similar to the fusion normalization layer RN. It combines its own instance normalization IN and the accumulated parameters corresponding to the batch normalization BN update during the training process of other source domains for normalization, thereby obtaining the ability to transfer feature extraction capabilities to distribution unknown inputs, thereby improving the generalization performance of the backbone network, extracting high-accuracy and high-precision image features, and improving the accuracy of unsupervised domain generalization person re-identification.

[0037] The first target feature vector and the second target feature vector are combined to obtain the recognition feature vector of the image to be recognized. Then, the recognition feature vector is processed with the target feature vector in the target domain for similarity measurement to obtain the recognition result of the target pedestrian corresponding to the image to be recognized. Therefore, through this recognition method, image datasets compatible with different source domains are realized, the generalization ability of pedestrian re-recognition technology is improved, the accuracy of pedestrian re-recognition is enhanced, and the efficiency of pedestrian re-recognition is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Also, throughout the accompanying drawings, the same reference figures are used to represent the same components. In the drawings:

[0039] Figure 1 A schematic diagram showing the steps of a pedestrian re-identification method in an embodiment of the present invention is shown;

[0040] Figure 2 A schematic diagram of the structure of the unsupervised domain generalized person re-identification training process based on the normalization layer improvement in an embodiment of the present invention is shown;

[0041] Figure 3 A schematic diagram of the structure of the ReNorm model framework of the pedestrian re-identification method in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0042] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0043] Embodiment 1

[0044] The first embodiment of the present invention provides a pedestrian re-identification method, such as Figure 1 As shown, including:

[0045] S101, obtaining an image to be recognized and a target recognition model, wherein the target recognition model includes a first target backbone network having L layers of fusion normalization layers and a second target backbone network having L layers of simulation normalization layers, the L layers of fusion normalization layers correspond to the L layers of simulation normalization layers one by one, and L≥1;

[0046] S102, firstly performing fusion normalized feature extraction on the image to be identified through the first target backbone network to obtain a first target feature vector of the image to be identified, and then performing simulated normalized feature extraction on the image to be identified through the second target backbone network to obtain a second target feature vector of the image to be identified;

[0047] S103, obtaining a recognition feature vector of the image to be recognized according to the first target feature vector and the second target feature vector;

[0048] S104, performing similarity measurement processing on the recognition feature vector and the target feature vector in the target domain to obtain a recognition result of the target pedestrian corresponding to the image to be recognized, wherein the target domain is a data set containing multiple target pedestrians.

[0049] It should be noted that the recognition model of this embodiment is a model capable of pedestrian re-identification, and the target recognition model is a trained recognition model. The first backbone network of the trained recognition model is the first target backbone network of the target recognition model. The second backbone network of the trained recognition model is the second target backbone network of the target recognition model. The data set of each target pedestrian in the target domain is the target feature vector of the target pedestrian.

[0050] Based on the pedestrian re-identification method of this embodiment, the first target backbone network with a fusion normalization layer RN (RemixNormalization) is used to perform fusion normalization feature extraction on the image to be identified, and the first target feature vector of the image to be identified is extracted. Then, the second target backbone network with an simulation normalization layer EN (Emulation Normalization) is used to perform simulation normalization feature extraction on the image to be identified, and the second target feature vector of the image to be identified is extracted. Among them, the fusion normalization layer RN (Remix Normalization) combines the existing instance normalization IN (Instance Normalization) and batch normalization BN (Batch Normalization) to normalize the statistical parameters. Batch normalization BN can extract discriminative features from the image. Instance normalization IN can eliminate the characteristics of domain information to a certain extent. The simulated normalization layer EN adopts a structure similar to the fusion normalization layer RN. It combines its own instance normalization IN and the accumulated parameters corresponding to the batch normalization BN update during the training process of other source domains for normalization, thereby obtaining the ability to transfer feature extraction capabilities to distribution unknown inputs, thereby improving the generalization performance of the backbone network, extracting high-accuracy and high-precision image features, and improving the accuracy of unsupervised domain generalization person re-identification.

[0051] The recognition feature vector of the image to be recognized is obtained by combining the first target feature vector and the second target feature vector. Then, the recognition feature vector is processed with the target feature vector in the target domain for similarity measurement to obtain the recognition result of the target pedestrian corresponding to the image to be recognized. The first target backbone network ensures that the model has a high recognition ability on the source domain data set. At the same time, the second target backbone network actively allows the model to learn how to transfer the recognition ability learned from the source domain training process with a known distribution to the target domain with an unknown distribution. Therefore, through this recognition method, an image data set that is compatible with different source domains is realized, the generalization ability of pedestrian re-identification technology is improved, the accuracy of pedestrian re-identification is enhanced, and the efficiency of pedestrian re-identification is improved.

[0052] Next, combine Figure 1 The specific implementation steps of the pedestrian re-identification method provided in this embodiment are introduced in detail:

[0053] Before executing step S101, that is, before obtaining the target recognition model, it is necessary to construct an initial recognition model, and then perform model training on the initial recognition model to obtain the target recognition model.

[0054] Model construction phase: Before training begins, the ReNorm model is selected as the recognition model of this embodiment, and the ResNet50 network structure is used as the model baseline of the backbone network. The ResNet50 model baseline contains 4 residual blocks, and there are a total of 48 BN layers in the 4 residual blocks as normalization layers. The unsupervised domain generalized pedestrian re-identification method based on the improvement of the normalization layer combines instance normalization IN and traditional batch normalization BN in various ways to ensure that the model can have both feature extraction performance and generalization performance.

[0055] like Figure 2As shown in part (a) of the figure, (a) is a schematic diagram of the training process of the existing unsupervised domain generalized pedestrian re-identification based on the normalization layer improvement. During the training of the improved ResNet50 model, the data of different source domains D1, D2, and D3 are input into the backbone network. The yellow area in the figure represents the operation process of the source domain D1, the red area in the figure represents the operation process of the source domain D2, and the blue area in the figure represents the operation process of the source domain D3. After the convolution layer, the improved normalization layer IN-BN1, IN-BN2, IN-BN3 and the activation function in the backbone network are processed, the output features are sent to the classifier of the improved ResNet50 model, and the classifier of the improved ResNet50 model outputs the classification result to complete the forward propagation. Afterwards, the loss function of the improved ResNet50 model calculates the loss according to the classification result output by the classifier of the improved ResNet50 model, and sends the loss to the backbone network to complete the backward propagation. The backbone network adjusts the internal parameters according to the received loss to complete an iterative training. In each iterative training, the number of forward propagations is 1, and only the first forward propagation is performed.

[0056] Based on the improved ResNet50 model, the ReNorm model framework of this recognition model will uniformly replace and modify these 48 normalization layers. This embodiment proposes two new complementary normalization layers: fusion normalization layer RN and simulation normalization layer EN. The ReNorm framework of this recognition model includes a first backbone network with L layers of fusion normalization layers RN and a second backbone network with L layers of simulation normalization layers EN. In essence, the first backbone network is a backbone network in which all batch normalization processing in the ResNet50 network is replaced with a fusion normalization layer RN formed by fusion normalization feature extraction, and the second backbone network is a backbone network in which all batch normalization processing in the ResNet50 network is replaced with a simulation normalization layer EN formed by simulation normalization feature extraction.

[0057] It should be noted that the number L of fusion normalization layers and simulation normalization layers in this embodiment can be set according to actual needs, for example, to 12 layers, 24 layers, 48 ​​layers or 96 layers. The following model is explained with a structure in which L is 48. Moreover, L layers of fusion normalization layers correspond one-to-one to L layers of simulation normalization layers, indicating that if there is one fusion normalization layer in the first backbone network, there is one simulation normalization layer corresponding to the fusion normalization layer in the second backbone network. In each iterative training of this recognition model, the processing process of each fusion normalization layer in the first backbone network is consistent, and the processing process of each simulation normalization layer in the second backbone network is consistent.

[0058] The implementation details of the ReNorm framework of this recognition model are described in combination with the training process of this recognition model, as follows:

[0059] After building this recognition model, the ReNorm backbone network is initialized with the RseNet50 model weights pre-trained on a certain dataset (such as the ImageNet dataset), and then the domain freezing DF (DomainFrozen) operation is performed to set the affine parameters of all the two normalization layers, the fusion normalization layer RN and the simulated normalization layer EN, to a non-updated state to avoid errors caused by overfitting of the normalization layer of this recognition model. In addition, the domain freezing DF operation is used to: the ReNorm framework will be initialized with the pre-trained ResNet50 weights before training begins. After initialization, the affine parameters of all RN normalization layers and EN normalization layers in the residual block are directly set to be non-updatable. Specifically, the γ parameters that need to be initialized in RN RN and β RN The corresponding position from the above weights is set to non-updatable after initialization. EN and β EN Directly set them uniformly to γ EN =1 and β EN =0.

[0060] In the model training phase: first obtain K source domains and an initial recognition model. The initial recognition model includes a first backbone network with L layers of fusion normalization layers and a second backbone network with L layers of simulation normalization layers. One source domain is a training image set of target pedestrians collected from the same camera network, K≥2. Then, the initial recognition model is iteratively trained through K source domains until the number of iterative training meets the iteration number threshold. The recognition model that meets the iteration number threshold is determined as the target recognition model. The iteration number threshold can be set according to actual needs.

[0061] In each iterative training of the initial recognition model, the affine parameters of the normalization layer of the initial recognition model are controlled to be in a non-updated state, and the affine parameters include: the scaling parameter γ and the translation parameter β of the normalization layer in the initial recognition model. The first backbone network is used to perform the first forward propagation processing on the input data corresponding to each iterative training to obtain the first prediction result of the input data, and the first loss data of the first forward propagation is obtained based on the first prediction result. The input data is a preset number of training images in a source domain of the K source domains under each iterative training.

[0062] It should be explained that, during the training process, the input data selected for each iterative training comes from one source domain randomly selected from the K source domains, so as to prevent multiple source domains from updating the recognition model together, causing the model to be biased towards the source domain that is easier to fit, and to avoid reducing the accuracy and generalization performance of the recognition model. The preset number can be set according to actual needs. For example, assuming that there are two source domains, each source domain has 128 training images, and each iterative training has 64 training images, then the order of iterative training input data can be: the input data corresponding to the first iterative training is the training images numbered 1-64 of the first source domain, the input data corresponding to the second iterative training is the training images numbered 65-128 of the first source domain, the input data corresponding to the third iterative training is the training images numbered 1-64 of the second source domain, and the input data corresponding to the fourth iterative training is the training images numbered 65-128 of the second source domain, and the iterative training is repeated in this way. To simplify the description, this embodiment describes the input data corresponding to each iterative training as one source domain, that is, it is assumed that the number of training images in one source domain is the number of input data corresponding to each iterative training of this recognition model.

[0063] The second backbone network then performs a second forward propagation process on the input data to obtain a second prediction result of the input data, and obtains second loss data of the second forward propagation based on the second prediction result. Then, the total loss data of the input data is obtained based on the first loss data and the second loss data, and based on the total loss data, the learnable parameters in the initial recognition model are updated through back propagation processing.

[0064] Specifically, if Figure 2 As shown in part (b) of FIG. 1 , (b) is a schematic diagram of the unsupervised domain generalized person re-identification training process based on the ReNorm framework structure of this embodiment. This ReNorm model framework only modifies and replaces the normalization layers in the four residual blocks, and integrates the above two complementary normalization layers through two forward propagations. In each iterative training, two forward propagations and one back propagation will be experienced. Figure 2 As shown in part (c) of the figure, (c) is a schematic diagram of the training details of RN and EN. In each iterative training, in the first forward propagation, the training data will only pass through the fusion normalization layer RN to ensure the discriminative feature extraction ability of the weight matrix. While the training data is transmitted in the fusion normalization layer RN, the statistical parameters of the batch normalization layer BN of the training data will be stored in the corresponding buffer module Buffer through the exponential moving average EMA. For example, in each layer of the fusion normalization layer RN, the buffer modules corresponding to the three source domains D1, D2, and D3 are Buffer 1 , Buffer 2 , Buffer 3 , that is, the buffer module corresponding to D1 is Buffer1 After passing through the first backbone network, the training data will be sent to the classifier corresponding to the fusion normalization layer RN (i.e., the first classifier) ​​and the corresponding loss will be calculated. Then the same training data will be sent to the ReNorm model framework for the second forward propagation, at which time the current input will be regarded as a simulated target domain of a randomly selected remaining source domain. For example Figure 2 and Figure 3 In the example, when D2 is randomly selected from D2 and D3, the training data of source domain D1 is input into RN. When passing through EN, the buffer corresponding to source domain D2 will be used. 2 . The training data will only pass through EN. At this time, EN is a simulation of the RN reasoning process, which will guide the weight matrix to learn how to transfer the feature extraction capability brought by RN to the unseen target domain. After passing through the second backbone network, the training data will be sent to the classifier (second classifier) ​​corresponding to the simulated normalization layer EN and calculate the corresponding loss. At the same time, the domain freezing DF is introduced to further obtain the weight matrix with domain invariance by freezing the update operation of all affine parameters in RN and EN.

[0065] The first forward propagation process: The first backbone network performs fusion normalization feature extraction on the input data to obtain the first feature vector. The first classifier then performs classification processing on the first feature vector to obtain the first prediction result. The process of fusion normalization feature extraction in each fusion normalization layer is as follows:

[0066]

[0067] like Figure 3 The gray part is shown as Figure 3 The upper part of the feature map for the input training image N, C, H and W represent the number of batch samples, number of channels, height and width of the feature map input to the fusion normalization layer RN, respectively. i∈(1,2,...,K) represents that the feature map is taken from the i-th one of the K source domains, that is, i is the ordinal variable of the source domain. Represents the input data for fused normalized feature extraction. express The output data after fusion normalization feature extraction, that is, the first feature vector of each fusion normalization layer, i∈(1,2,...,K), K is the total number of source domains. γ RN represents the non-learnable scaling parameter in the fusion normalization feature extraction, β RN represents the non-learnable translation parameter in the fusion normalization feature extraction, and ∈ is a constant to prevent the denominator from being zero. Corresponding to input The mean and variance of the statistical parameters after instance normalization, Corresponding to input Mean and variance of statistical parameters after batch normalization.

[0068] In each fusion normalization layer, obtain The method is:

[0069]

[0070]

[0071] Among them, x nchw Represents input A point in . n,h,w are variables in the range of N, H and W respectively.

[0072] During the first forward propagation process, in each fusion normalization layer, the mean and variance of the accumulated parameters in the buffer module corresponding to the input data are also updated based on the mean and variance after batch normalization. Specifically, due to the special mechanism of BN, this embodiment sets a total of K cache modules Buffer corresponding to the number of source domains in each fusion normalization layer RN. 1,2,...,K To store the accumulated parameters in the first forward propagation When input Each time passing through RN, the corresponding Buffer i The corresponding

[0073]

[0074] in, is the mean of the cumulative parameters, is the cumulative variance of the parameters, and λ is the update weight set, usually set to 0.9.

[0075] For example, the first backbone network of this model has 48 layers of fusion normalization layers RN, and the number of source domains is K. In each iterative training process of this model, the first backbone network performs the first forward propagation processing on the input data of a certain source domain. In the first forward propagation processing, in the first layer RN, the input data is subjected to fusion normalization feature extraction to obtain In the first layer RN, there are K buffer modules Buffer 1,2,...,K , Buffer 1,2,...,K Store the accumulated parameters of the source domain corresponding to the first forward propagation, that is, in the first layer RN, Buffer 1 Store the corresponding cumulative parameters of the source domain D1 Buffer 2 Store the corresponding cumulative parameters of the source domain D2 And so on. In the second layer RN, the input data is fused and normalized to extract features, and the obtained In the second layer RN, there are still K buffer modules Buffer 1,2,...,K , Buffer 1,2,...,K Store the accumulated parameters of the source domain corresponding to the first forward propagation, that is, in the second layer RN, Buffer 1 Store the corresponding cumulative parameters of the source domain D1 Buffer 2 Store the corresponding cumulative parameters of the source domain D2 And so on. In this way, each layer of RN extracts the fused normalized features of the input data and obtains Then, each layer of RN sets K buffer modules 1,2,...,K To store the accumulated parameters in the first forward propagation

[0076] In the first forward propagation, the training data first passes through the first backbone network composed of RN in the residual block to obtain the first eigenvector, that is, Then, the first feature vector is classified by the classifier corresponding to the fusion normalization layer (i.e., the first classifier) ​​to obtain the first prediction result, i.e., the prediction probability p. The loss function of the first forward propagation (i.e., the first loss data) includes: the ternary loss of the first backbone network and the classification loss of the first classifier The ternary loss of the first backbone network and the classification loss of the first classifier The calculation method is as follows:

[0077]

[0078] Among them, n∈{1,2,3,…,N}, id∈{1,2,3,…,ID}, ID is the number of pedestrian identities in the i-th source domain, N is the number of samples in the i-th source domain, y is the real pedestrian identity label, p k represents the probability that the model predicts that the sample belongs to the pedestrian identity label k, ε 0 is a constant used for label smoothing, q k It is a "soft label" after label smoothing.

[0079]

[0080] Among them, ||·|| represents L 2 norm, and Respectively represent the features f of the nth input sample nFor positive and negative samples in a batch, margin is a boundary constant.

[0081] Classification loss corresponding to RN and ternary loss The classification loss corresponding to EN and ternary loss is calculated in the same way.

[0082] Based on the fusion normalization feature extraction process, the calculation method of the fusion normalization layer RN reasoning process (i.e., the RN recognition process) is as follows:

[0083]

[0084] At this time, the statistical parameters of IN required by the fusion normalization layer RN and Still comes from the input target domain feature map itself, but the statistical parameters of BN required by the fusion normalization layer RN come from the accumulated data in the source domain training process. and Since the two come from different data distributions, there is a large inconsistency. As a result, the input target domain feature map cannot be correctly Normalization is performed, which results in abnormal distribution of output data, and the recognition model does not have the ability to actively deal with this situation. As a result, the performance of the model trained in the source domain is degraded in the unknown target domain.

[0085] Therefore, by introducing EN to simulate the RN reasoning process, the weight matrix is ​​guided to learn how to transfer the feature extraction capability brought by RN to the unseen target domain.

[0086] The second forward propagation process: the input data is subjected to simulated normalization feature extraction through the simulated normalization layer to obtain the second feature vector. The second feature vector is then classified through the second classifier to obtain the second prediction result. The process of simulated normalization feature extraction in each simulated normalization layer is as follows:

[0087]

[0088] like Figure 3 As shown in the colored part, Figure 3 The lower part of the feature map for the input training image Where N, C, H and W represent the number of batch samples, number of channels, height and width of the feature map input to the simulated normalization layer EN, i represents that the current data is taken from the i-th source domain in the total K source domains, j is the j-th source domain in the K source domains, i, j∈(1,2,...,K), and i≠j. j is the ordinal variable of the source domain, j∈(1,2,...,K) and i≠j, represents the input data for simulated normalized feature extraction, Represents input The output data after simulated normalization feature extraction, that is, the second eigenvector of each simulated normalization layer. EN represents the non-learnable scaling parameter in simulated normalization feature extraction, β EN represents the non-learnable translation parameter in simulated normalization feature extraction. yes The mean and variance of the statistical parameters after instance normalization, It comes from the jth buffer module Buffer j The mean and variance of the cumulative parameters stored in a j , b j Respectively represent Buffer j The two learnable parameters are used to adaptively learn the statistical parameters and cumulative parameters in the corresponding instance normalization processing.

[0089] It needs to be explained that the input source domain D i A source domain D is randomly selected from the remaining K-1 source domains j , D i Will be D j In the second forward propagation, all normalization layers in the residual block of the backbone network will be replaced by EN, i.e., the second backbone network. In each layer of EN, there will be i The input feature map provides the required IN statistical parameters And by the selected D j Provides the statistical parameters of BN accumulation required by EN For example, the second backbone network of this model has 48 layers of simulated normalization layers EN, and the number of source domains is K. In each iterative training process of this model, the source domain D1 is forward propagated for the first time through the first backbone network, and D1 is forward propagated for the second time through the second backbone network. During the second forward propagation process, D2 is randomly selected as the simulated target domain of D1. In the first layer EN, simulated normalization feature extraction is performed on D1, that is, IN statistical parameters are provided by D1. And D2 provides the statistical parameters of BN accumulation In the second layer EN, simulated normalized feature extraction is performed on D1, that is, IN statistical parameters are provided by D1 And D2 provides the statistical parameters of BN accumulation Processing, and so on.

[0090] It can be seen that in the simulated normalization layer EN, a network structure similar to that of the fusion normalization layer RN is adopted. And with the help of the distribution differences between the K training source domains, the input source domain i is regarded as the simulated target domain of a randomly selected remaining source domain j. Similar to the reasoning process of the fusion normalization layer RN, the statistical parameters of IN at this time are and Still comes from the input feature map The statistical parameters of BN also come from the accumulated data in the source domain j during the training process. and The distributions of the two sets of statistical parameters are inconsistent. Therefore, with the help of the simulated normalization layer EN, the recognition model actively obtains the ability to transfer feature extraction capabilities to distribution unknown inputs, thereby improving the generalization performance of the weight matrix and improving the accuracy of unsupervised domain generalization pedestrian re-identification.

[0091] In the second forward propagation, the training data first passes through the second backbone network composed of EN in the residual block to obtain the second eigenvector, that is, Then, the second feature vector is classified by simulating the classifier corresponding to the normalization layer (i.e., the second classifier) ​​to obtain the second prediction result. The loss function of the second forward propagation (i.e., the second loss data) includes: the ternary loss of the second backbone network and the classification loss of the second classifier

[0092] According to the first loss data and the second loss data, the total loss data of the input data is obtained, and based on the total loss data, the learnable parameters in the initial recognition model are updated through back propagation processing. The total loss data of each iterative training From the total loss data L total Calculate the gradients required for back-propagation update and perform back-propagation to update the learnable parameters inside the ResNet50 model.

[0093] This embodiment is based on the proposed renormalization ReNorm network, and uses two complementary normalization operations to perform different normalization processing on the same weight matrix through two forward propagations and one back propagation. In the first forward propagation process, the newly proposed fusion normalization layer RN method is used to replace all batch normalization BN layers in the residual block of the original backbone network ResNet50. RN obtains the statistical parameters required for standardization by directly adding the statistical parameters of the instance normalization IN and batch normalization BN of the input feature map. Subsequently, in the second forward propagation, RN is completely replaced by the simulated normalization layer EN, which simulates the reasoning process of RN, and uses the statistical parameters of IN of the input feature map and the cumulative parameters of BN provided by a randomly selected residual source domain to obtain the parameters required for standardization. Furthermore, the affine parameters of RN and EN are updated during the domain freezing DF training process to reduce the impact of this parameter on the model performance during training and testing.

[0094] Combination Figure 3 Briefly explain the process of the single-layer fusion normalization layer RN and the single-layer simulation normalization layer EN corresponding to the RN in each iterative training (the process of the single-layer fusion normalization layer RN and the single-layer simulation normalization layer EN corresponding to the RN is consistent with the process of other layers RN and EN). The source domain D1 passes through the first backbone network of RN to obtain the first feature vector of D1. While transmitting in the fusion normalization layer RN, the statistical parameters of the batch normalization layer BN of D1 are stored in the corresponding buffer module Buffer through the exponential moving average EMA. 1 After passing through the first backbone network, D1 will be sent to the fusion normalization layer RN classifier and the corresponding loss will be calculated. Then D1 will be sent to the ReNorm model framework for the second forward propagation. When the source domain D1 is input to RN and then passes through EN, Buffer will be used. 2 The corresponding source domain D2. D1 will only pass through EN. At this time, EN is a simulation of the RN reasoning process, which will guide the weight matrix to learn how to transfer the feature extraction capability brought by RN to the unseen target domain. After passing through the second backbone network, D1 will be sent to the simulated normalization layer EN classifier and calculate the corresponding loss. At the same time, the domain freezing DF is introduced to further obtain the weight matrix with domain invariance by freezing the update operation of all affine parameters in RN and EN.

[0095] After obtaining the target recognition model, step S101 is executed to obtain the image to be recognized and the target recognition model, wherein the target recognition model includes a first target backbone network having L layers of fusion normalization layers and a second target backbone network having L layers of simulation normalization layers, and the L layers of fusion normalization layers correspond one to one to the L layers of simulation normalization layers, and L≥1. Specifically, after the initial recognition model is trained on K source domains, it will be directly tested and used on an unseen target domain. At this time, only the backbone network of the model will be retained. The image with recognition is input to the target recognition model for direct recognition on an unseen target domain.

[0096] Next, execute step S102, first perform fusion normalized feature extraction on the image to be identified through the first target backbone network to obtain the first target feature vector of the image to be identified, and then perform simulated normalized feature extraction on the image to be identified through the second target backbone network to obtain the second target feature vector of the image to be identified.

[0097] Specifically, the image to be recognized is input into the backbone network of RN (i.e., the first target backbone network) K times, and each time the input feature map (i.e., the image to be recognized) is normalized using the model parameters corresponding to the K source domains in turn, and K feature vectors of the image to be recognized in the first forward propagation are obtained in turn, i.e., the first target feature vector. Subsequently, the image to be recognized is input into the backbone network of EN (i.e., the second target backbone network) K times, and each time the Buffer 1 , Buffer 2 , ..., Buffer K The information in (a 1 , b 1 ), (a 2 , b 2 ),...,(a K , b K ) A total of K groups of EN internal balance parameters are input into the feature map for normalization operation, and K feature vectors of the image to be identified in the second forward propagation are obtained in turn, that is, the second target feature vector.

[0098] Then, step S103 is executed to obtain the recognition feature vector of the image to be recognized based on the first target feature vector and the second target feature vector. Specifically, the image to be recognized will complete 2*K forward propagations through the backbone network respectively, and the recognition feature vector representing the image to be recognized will be output after the average of all the results is calculated, that is, the first target feature vector and the second target feature vector are averaged, and the average value is taken as the feature vector to be recognized.

[0099] S104, performing similarity measurement processing on the recognition feature vector and the target feature vector in the target domain to obtain the recognition result of the target pedestrian corresponding to the image to be recognized, wherein the target domain is a data set containing multiple target pedestrians. Specifically, performing a similarity measurement operation such as Euclidean distance similarity or cosine similarity on the recognition feature vector and the target feature vector to obtain the recognition result of the target pedestrian corresponding to the image to be recognized.

[0100] In this embodiment, K source domains are used to train the renormalized ReNorm model to obtain a target recognition model. During the training of the target recognition model, the ReNorm model replaces all batch normalization BN in the backbone network with a fusion normalization layer RN or an analog normalization layer EN. Two forward propagations are required for feature extraction during the model training process. The fusion normalization layer RN passed through by the first forward propagation combines the existing instance normalization IN and batch normalization BN statistical parameters for normalization. The second forward propagation adopts a structure similar to the fusion normalization layer RN, and uses the distribution differences between the K training source domains, combined with the instance normalization IN of the current source domain and the cumulative parameters corresponding to the batch normalization BN update from other source domains during training to normalize, thereby obtaining the ability to transfer feature extraction capabilities to distribution unknown inputs, thereby improving the generalization performance of the backbone network, thereby improving the accuracy of unsupervised domain generalization pedestrian re-identification. Based on the target recognition model of this embodiment, it is possible to achieve compatibility with image data sets in different source domains, improve the generalization ability of pedestrian re-identification technology, enhance the accuracy of pedestrian re-identification, and thus improve the efficiency of pedestrian re-identification.

[0101] The recognition method of this embodiment is cross-domain trained on a cross-domain unsupervised test dataset in two experimental scenarios, three experimental settings, and a total of 9 pedestrian re-identification datasets, and its effectiveness is verified through experiments. The specific experimental settings include:

[0102] Small dataset target domain scenario: Four large datasets are used as source domains, namely Market1501 (M), CUHK02 (C2), CUHK03 (C3), and CUHK-SYSU (CS). Four small datasets, PRID, GRID, VIPeR, and iLIDs, are used as target domains. As shown in Table 1, the comparison with other unsupervised domain generalization person re-identification methods in the small dataset target domain scenario is as follows:

[0103] Table 1

[0104]

[0105] As can be seen from Table 1, in terms of average performance, this method is at least 2 percentage points higher than the previous method in terms of mAP and at least 3.3 percentage points higher than the previous method in terms of Rank-1. And among the four test target domains, this method achieved the best performance on three target domain datasets. The experiment fully demonstrates that this method has greatly improved the generalization performance of the model.

[0106] Large dataset target domain scenario: It involves 4 large datasets: Market1501(M), MSMT17(MS), CUHK03(C3) and CUHK-SYSU(CS). The experiment was conducted according to the "leave one out" method. CUHK03(C3), MSMT17(MS) and Market1501(M) were used as target domains, and the remaining datasets were used as source domains. In the current scenario, it is further divided into two experimental settings, the difference is: Experimental setting 1: The training process only uses the training set of the source domain dataset. Experimental setting 2: The training process will use all images of the dataset, including the training set and the test set. As shown in Table 2, it is compared with other unsupervised domain generalization pedestrian re-identification methods in the large dataset target domain scenario:

[0107] Table 2

[0108]

[0109] As can be seen from Table 2: In the two experimental settings, out of a total of 16 evaluation indicators, this method achieved the best performance in 14 indicators. In terms of average performance, in the experimental setting 1, this method is at least 2 percentage points higher than the previous method in the mAP indicator, and at least 4.1 percentage points higher than the previous method in the Rank-1 indicator. In the experimental setting 2, this method is at least 1.7 percentage points higher than the previous method in the mAP indicator, and at least 2.9 percentage points higher than the previous method in the Rank-1 indicator. The above experiments further prove the effectiveness of this method in the target domain scenario of large datasets.

[0110] The experiment finally showed that under the three experimental settings, the recognition method proposed in this embodiment achieved the SOTA (highest performance level) of the average indicator. And in a total of 10 groups of experiments, a total of 20 corresponding evaluation indicators, 16 indicators achieved SOTA. The experiment shows the effectiveness of this recognition method. As shown in Tables 1 and 2, the method ReNorm of the present invention achieved SOTA in both mAP and Rank-1 indicators under the experimental settings of the target domains PRID, GRID, and VIPeR, and the two indicators were 2.0% and 3.3% higher than the previous SOTA in terms of average indicators. The experiment proves the effectiveness of this recognition method.

[0111] Embodiment 2

[0112] Based on the same inventive concept, the second embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the above-mentioned pedestrian re-identification methods when executing the program.

[0113] Embodiment 3

[0114] Based on the same inventive concept, the third embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of any method of the pedestrian re-identification method described in the first embodiment above are implemented.

[0115] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0117] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0118] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0119] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0120] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A pedestrian re-identification method, characterized in that: include: Acquire an image to be recognized and a target recognition model, wherein the target recognition model includes a first target backbone network having L layers of fusion normalization layers and a second target backbone network having L layers of simulation normalization layers, wherein the L layers of fusion normalization layers correspond one to one to the L layers of simulation normalization layers, and L≥1; Firstly, the first target backbone network is used to perform fusion normalized feature extraction on the image to be identified to obtain a first target feature vector of the image to be identified, and then the second target backbone network is used to perform simulated normalized feature extraction on the image to be identified to obtain a second target feature vector of the image to be identified; Obtaining an identification feature vector of the image to be identified according to the first target feature vector and the second target feature vector; The recognition feature vector is processed with a target feature vector in a target domain to measure similarity, so as to obtain a recognition result of a target pedestrian corresponding to the image to be recognized, wherein the target domain is a data set containing multiple target pedestrians.

2. The method according to claim 1, characterized in that Before obtaining the target recognition model, the following steps are included: Obtain K source domains and an initial recognition model, wherein the initial recognition model includes a first backbone network having the L-layer fusion normalization layer and a second backbone network having the L-layer simulation normalization layer, one source domain is a training image set of a target pedestrian collected from the same camera network, and K≥2; Iteratively training the initial recognition model through the K source domains until the number of iterative trainings meets an iteration number threshold, and determining the recognition model that meets the iteration number threshold as the target recognition model; wherein, in each iterative training of the initial recognition model, the affine parameters of the normalization layer of the initial recognition model are controlled to be in a non-updated state, the first backbone network is used to perform a first forward propagation process on the input data corresponding to each iterative training to obtain a first prediction result of the input data, and first loss data of the first forward propagation is obtained based on the first prediction result, and the input data is a preset number of training images in a source domain of the K source domains under each iterative training; Performing a second forward propagation process on the input data through the second backbone network to obtain a second prediction result of the input data, and obtaining second loss data of the second forward propagation based on the second prediction result; According to the first loss data and the second loss data, the total loss data of the input data is obtained, and based on the total loss data, the learnable parameters in the initial recognition model are updated through back propagation processing.

3. The method according to claim 2, characterized in that The first backbone network is used to perform a first forward propagation process on the input data corresponding to each iterative training to obtain a first prediction result of the input data, including: Performing fusion normalization feature extraction on the input data through the first backbone network to obtain a first feature vector, and then performing classification processing on the first feature vector through a first classifier to obtain the first prediction result; The process of extracting the fused normalized features in each fused normalized layer is as follows: i is the ordinal variable of the source domain, represents the input data of the fused normalized feature extraction, express The output data after fusion normalization feature extraction, that is, the first feature vector of each fusion normalization layer, i∈(1,2,...,K), K is the total number of source domains, γ RN represents the non-learnable scaling parameter in the fused normalized feature extraction, β RN represents the non-learnable translation parameter in the fusion normalization feature extraction, ∈ is a constant to prevent the denominator from being 0, Corresponding to input The mean and variance of the statistical parameters after instance normalization, Corresponding to input Mean and variance of statistical parameters after batch normalization.

4. The method according to claim 3, characterized in that After the first backbone network performs fusion normalization feature extraction on the input data, the method further includes: In each fusion normalization layer, based on the mean after the batch normalization processing and the variance after the batch normalization processing, the mean and variance of the accumulated parameters in the buffer module corresponding to the input data are updated, as shown in the following expression: in, is the mean of the cumulative parameter, is the variance of the accumulated parameters, and λ is the update weight set.

5. The method according to claim 4, characterized in that The performing a second forward propagation process on the input data through the second backbone network to obtain a second prediction result of the input data includes: Performing simulated normalized feature extraction on the input data set through the second backbone network to obtain a second feature vector, and then performing classification processing on the second feature vector through a second classifier to obtain the second prediction result; The process of extracting the simulated normalized features in each simulated normalized layer is as follows: j is the ordinal variable of the source domain, j∈(1,2,...,K) and i≠j, represents the input data for simulated normalized feature extraction, Represents input The output data after simulated normalization feature extraction, i.e., the second feature vector of each simulated normalization layer, γ EN represents the non-learnable scaling parameter in simulated normalization feature extraction, β EN represents the non-learnable translation parameter in simulated normalization feature extraction, yes The mean and variance of the statistical parameters after instance normalization, It comes from the jth buffer module Buffer j The mean and variance of the cumulative parameters stored in a j , b j Respectively represent Buffer j The two learnable parameters are used to adaptively learn the statistical parameters and cumulative parameters in the corresponding instance normalization processing.

6. The method according to claim 2, characterized in that The first loss data includes: the ternary loss of the first backbone network and the classification loss of the first classifier The second loss data includes: the ternary loss of the second backbone network and the classification loss of the second classifier The total loss data 7. The method according to claim 2, characterized in that The affine parameters include: a scaling parameter γ and a translation parameter β of the normalization layer in the initial recognition model.

8. The method according to claim 2, characterized in that The target recognition model and the initial recognition model are both ReNorm models. The first backbone network is a backbone network in which all batch normalization processing in the ResNet50 network is replaced by a fused normalization layer formed by fused normalization feature extraction. The second backbone network is a backbone network in which all batch normalization processing in the ResNet50 network is replaced by a simulated normalization layer formed by simulated normalization feature extraction.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method steps according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method steps described in any one of claims 1 to 8 are implemented.