Double-branch neural network pedestrian re-identification method based on multi-granularity feature fusion

Abstract

The invention provides a double-branch neural network pedestrian re-identification method based on multi-granularity feature fusion, and the method mainly aims at pedestrian image multi-granularity feature fusion, employs a global branch and feature fusion branch parallel calculation mode, and outputs global features and fusion features. Finally, the features of the global branches and the fusion branches are fused as final features, and the features are more accurate and rich. According to the method, shallow texture features and deep semantic features extracted from pedestrian images can be effectively fused, and target detail information is prevented from being lost; the global branch and the fusion branch are adopted for feature extraction, it is guaranteed that the extracted features have the maximum discriminability, the utilization rate of the algorithm to the image features is improved, and the extracted features can be fused with more feature information of the current target. The pedestrian re-identification accuracy is improved, and the interference of factors such as camera view change and background change on pedestrian re-identification is weakened.

Description

Technical Field

[0001] The present invention belongs to the technical field of pedestrian re-identification, and in particular relates to a double-branch neural network pedestrian re-identification method based on multi-granularity feature fusion. Background Art

[0002] Pedestrian re-identification is a technology that uses computer vision algorithms to search for pedestrians with the same identity in a multi-camera surveillance network with non-overlapping shooting areas. For most surveillance cameras, pedestrians in the picture are the main focus, and identifying pedestrians in the picture is a major task in the security field. As the main infrastructure for social public security, video surveillance systems have been widely used in public places such as roads, stations, shopping malls, campuses, and communities, playing an important role in public security investigations, illegal capture, and danger prediction.

[0003] In the task of pedestrian re-identification, the main characteristics of pedestrians include two aspects. One is the appearance characteristics of clothing, including the color and texture of clothing, head ornaments, hairstyles, glasses, hats, foot shape, shoe texture, objects and decorations carried by hands, etc. The other is the appearance characteristics of pedestrians, including the posture of pedestrians, gait when walking, sitting posture when riding a bicycle, and movements of the head and upper limbs. The appearance characteristics and appearance characteristics of pedestrians are the main basis for pedestrian identification. However, due to the complexity of actual monitoring scenes and the non-rigidity of pedestrian structures, the appearance characteristics and appearance characteristics of pedestrians in different domains are not unique, distinctive and stable. The appearance characteristics and appearance characteristics of the same pedestrian in different domains will also change greatly, resulting in the feature distances of different pedestrians being too close and the feature distances of the same pedestrian being distant, which affects the model's extraction of robust pedestrian feature descriptions and causes mismatches. How to accurately extract pedestrian identification features is an important difficulty in the task of pedestrian re-identification. Summary of the invention

[0004] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art, and proposes a dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion. For the multi-granularity feature fusion of pedestrian images, a global branch and a feature fusion branch are used in parallel calculation to output global features and fusion features. Finally, the features of the global branch and the fusion branch are fused as the final feature, which has more accurate and rich features. Thereby improving the accuracy of pedestrian re-identification and reducing the interference of factors such as camera field of view changes and background changes on pedestrian re-identification.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] The first aspect of the present invention provides a dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion, comprising the following steps:

[0007] Step 1: Obtain source and target domain datasets for person re-identification;

[0008] Step 2: Construct a parallel training architecture, which includes four feature extraction networks with the same network structure, which are used to generate hard labels and soft labels respectively. Each feature extraction network includes two branches: a global branch and a feature fusion branch. The source domain data is input into the parallel training architecture.

[0009] Step 3: The global branch uses ResNet-50 to learn the global feature c5 of each image in the source domain data. The step size of res_conv5x in ResNet-50 is 1, and the output size is [16, 8];

[0010] Step 4: The fusion branch uses a pyramid network to fuse the c2, c3, c4, and c5 features in the global branch ResNet-50, where c2 is the output result after the input passes through res2, with a size of [64, 32] and 256 channels, c3 is the result after c2 passes through res3, with a size of [32, 16] and 512 channels, c4 is the result after c3 passes through res4, with a size of [16, 8] and 1024 channels, and c5 is the result after c4 passes through res The result after s5 is of size [16,8], 2048 channels, and the outputs are p2, p3, and p4. P4 is obtained by adding c4 after upsampling and adjusting the channels of c5, with a size of [16,8], 256 channels. P3 is obtained by adding c3 after upsampling and adjusting the channels of p4, with a size of [32,16], 256 channels. P2 is obtained by adding c2 after upsampling and adjusting the channels of p3, with a size of [64,32], 256 channels.

[0011] Step 5: Resize p2, p3, and p4 to [16, 8] by downsampling to obtain p2', p3', and p4';

[0012] Step 6: Splice p2', p3', and p4' along the channel, compress the channel to 1024, and obtain c4_f;

[0013] Step 7: Input c4_f into the res_conv5x module to obtain c5_f. The parameters of the res_conv5x module are the same as those in the global branch.

[0014] Step 8: Use the loss function to constrain the entire network structure;

[0015] Step 9: Determine whether the number of iterations set during network training has been reached. If yes, execute step 10; if no, execute step 3;

[0016] Step 10: Fuse c5 and c5_f obtained from each pedestrian image to obtain the final feature feature_r for pedestrian re-identification;

[0017] Step 11: Use the distance measurement method to determine the identity of each pedestrian's feature feature_r and output the recognition result.

[0018] Furthermore, in step 8, the loss function is expressed by the following expression:

[0019]

[0020]

[0021] is the hard label identity loss in the feature extraction network, is the hard label triplet loss in the feature extraction network, is the soft label identity loss in the feature extraction network, is the soft label triplet loss in the feature extraction network, and is the weight coefficient, is the classification of the target domain under the four networks, They are the feature output of the main branch and the feature output of the fusion branch under the same network.

[0022] They represent the current feature, the most difficult positive sample, and the most difficult negative sample respectively. The most difficult positive sample represents the feature that is the most dissimilar to the current one in the current training batch data. The most difficult negative sample represents the feature that is the most similar to the current one in the current training batch data. express is the true weight matrix of the kth person, E (T) is the balance coefficient between the Tth iteration and the T-1th iteration of each model weight parameter.

[0023] Furthermore, in step 10, the fusion method is to take the maximum value of each dimension of c5 and c5_f as the final feature_r.

[0024] The second aspect of the present invention provides a dual-branch neural network pedestrian re-identification device based on multi-granularity feature fusion, comprising:

[0025] Data acquisition unit: used to obtain source domain and target domain data sets for person re-identification;

[0026] Construction unit: used to construct a parallel training architecture, which includes four feature extraction networks with the same network structure, which are used to generate hard labels and soft labels respectively. Each feature extraction network includes two branches: a global branch and a feature fusion branch, and the source domain data is input into the parallel training architecture;

[0027] The first processing unit: used for the global branch to use ResNet-50 to learn the global feature c5 of each image in the source domain data. The step size of res_conv5x in ResNet-50 is 1, and the output size is [16,8];

[0028] The second processing unit is used to fuse the c2, c3, c4, and c5 features in the global branch ResNet-50 using a pyramid network, where c2 is the output result after the input passes through res2, with a size of [64, 32] and 256 channels, c3 is the result after c2 passes through res3, with a size of [32, 16] and 512 channels, c4 is the result after c3 passes through res4, with a size of [16, 8] and 1024 channels, and c5 is the result after c4 passes through res4. The result after res5 is of size [16,8], 2048 channels, and the output is p2, p3, p4. P4 is obtained by adding c4 after upsampling and adjusting the channel of c5, with a size of [16,8], 256 channels. P3 is obtained by adding c3 after upsampling and adjusting the channel of p4, with a size of [32,16], 256 channels. P2 is obtained by adding c2 after upsampling and adjusting the channel of p3, with a size of [64,32], 256 channels.

[0029] The third processing unit is used to adjust the size of p2, p3, and p4 to [16,8] by downsampling to obtain p2', p3', and p4';

[0030] The fourth processing unit is used to splice p2', p3', and p4' along the channel, compress the channel to 1024, and obtain c4_f;

[0031] The fifth processing unit is used to input c4_f into the res_conv5x module to obtain c5_f. The parameters of the res_conv5x module are the same as those of the res_conv5x module in the global branch.

[0032] The sixth processing unit is used to constrain the entire network structure using a loss function;

[0033] The seventh processing unit is used to determine whether the number of iterations set during network training has been reached. If yes, the eighth processing unit is executed; if no, the first processing unit is executed;

[0034] The eighth processing unit: fuses c5 and c5_f obtained from each pedestrian image to obtain the feature feature_r used for pedestrian re-identification;

[0035] Result output unit: used to identify the pedestrian using a distance metric for each pedestrian's feature feature_r and output the recognition result.

[0036] A third aspect of the present invention provides an electronic device, comprising

[0037] at least one processor, and

[0038] at least one memory in communication with the processor, wherein:

[0039] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the above-mentioned dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion.

[0040] A fourth aspect of the present invention provides a non-volatile computer-readable storage medium, which, when the computer-executable instructions are executed by one or more processors, enables the processors to execute the above-mentioned dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion.

[0041] Compared with the prior art, the dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion described in the present invention has the following advantages:

[0042] The present invention proposes a dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion, which can effectively fuse shallow texture features and deep semantic features extracted from pedestrian images to avoid the loss of target detail information; by adopting global branches and fusion branches to extract features respectively, it is ensured that the extracted features have the maximum degree of discriminability, improves the algorithm's utilization of image features, and enables the extracted features to fuse more feature information of the current target; the present invention improves the accuracy of pedestrian re-identification and reduces the interference of factors such as camera field of view changes and background changes on pedestrian re-identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings: Figure 1 It is an overall flow chart of a double-branch neural network pedestrian re-identification method based on multi-granularity feature fusion of the present invention;

[0044] Figure 2 It is a schematic diagram of the training network framework of the present invention;

[0045] Figure 3 This is a schematic diagram of a dual-branch network for feature extraction of the present invention;

[0046] Figure 4 It is a schematic diagram of the global feature and fusion feature fusion method of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0048] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0049] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0050] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0051] Embodiment 1:

[0052] The pedestrian re-identification algorithm of the present invention realizes the multi-granularity feature fusion of pedestrian images, adopts the global branch and feature fusion branch parallel calculation mode, outputs the global feature and the fusion feature. Finally, the features of the global branch and the fusion branch are fused as the final feature. Figure 1 This is the overall flow chart of a dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion proposed in the present invention. The specific steps are as follows:

[0053] 1) Based on the mutual teaching parallel training architecture, unsupervised cross-domain pedestrian re-identification is trained. The architecture contains four feature extraction networks: NetA(θ1), NetB(θ2), MeanNetA(θ3), and MeanNetB(θ4). NetA(θ1) and NetB(θ2) are used to generate hard labels, which are used to determine uniqueness. MeanNetA(θ3) and MeanNetB(θ4) are used to generate soft labels, which are used to improve the accuracy of pseudo labels. The structure of each network is the same. MeanNetA and MeanNetB are used to supervise and correct NetA and NetB respectively. Pedestrian images are input into the four feature extraction networks respectively.

[0054] 2) The feature extraction network contains two branches, one is the global branch and the other is the feature fusion branch. The global branch uses ResNet-50 to learn the global feature c5 of pedestrian images. The step size of res_conv5x in ResNet-50 is 1 and the output size is [16,8].

[0055] 3) The fusion branch uses a pyramid network to fuse the c2, c3, c4, and c5 features in the global branch ResNet-50, where c2 is the output result after the input passes through res2, with a size of [64, 32] and 256 channels, c3 is the result after c2 passes through res3, with a size of [32, 16] and 512 channels, c4 is the result after c3 passes through res4, with a size of [16, 8] and 1024 channels, and c5 is the result after c4 passes through res5, with a size of [16, 8] and 2048 channels. The outputs are p2, p3, and p4. P4 is obtained by adding c5 to c4 after upsampling and adjusting the channel, with a size of [16, 8] and 256 channels. P3 is obtained by adding p4 to c3 after upsampling and adjusting the channel, with a size of [32, 16] and 256 channels. P2 is obtained by adding p3 to c2 after upsampling and adjusting the channel, with a size of [64, 32] and 256 channels.

[0056] 4) Adjust the size of p2, p3, and p4 to [16, 8] and the channel to 1024 by downsampling to obtain p2', p3', and p4'.

[0057] 5) Concatenate p2', p3', and p4' along the channel, and then compress the channel to 1024 to obtain c4_f.

[0058] 6) Input c4_f into the res_conv5x module to obtain c5_f. The parameters of the res_conv5x module are the same as those of the res_conv5x module in the global branch.

[0059] 7) Use the loss function L(θ1,θ2,θ3,θ4) to constrain the entire network structure. The network loss function is expressed by the following expression:

[0060]

[0061]

[0062] is the hard label identity loss in the feature extraction network, Hard label triplet loss in feature extraction network. is the soft label identity loss in the feature extraction network, Soft label triplet loss in feature extraction network. and is the weight coefficient. is the classification of the target domain under the four networks. They are the feature output of the global branch and the feature output of the fusion branch under the same network.

[0063] They represent the current feature, the most difficult positive sample and the most difficult negative sample respectively. The most difficult positive sample represents the feature that is the most dissimilar to the current one in the current training batch data. The most difficult negative sample represents the feature that is the most similar to the current one in the current training batch data. express is the true weight matrix of the kth person. E (T) is the balance coefficient between the Tth iteration and the T-1th iteration of each model weight parameter.

[0064] 8) Determine whether the number of iterations set during network training has been reached. If yes, execute step 9) If no, execute step 2)

[0065] 9) The c5 and c5_f obtained from each pedestrian image are fused to obtain the final feature feature_r for pedestrian re-identification. The specific fusion method is to take the maximum value of each dimension of c5 and c5_f as the final feature_r.

[0066] 10) The feature feature_r of each pedestrian is used to determine the identity of the pedestrian using a distance metric. The closer the distance, the greater the possibility that they are the same person, and the greater the distance, the greater the possibility that they are not the same person.

[0067] The dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion proposed in the present invention realizes the multi-granularity feature fusion of pedestrian images, and at the same time optimizes the extraction of global image features and fused image features by optimizing network hyperparameters, and finally encodes the global image features and fused features into representation features for pedestrian re-identification. Here, Resnet50 and IBN-Resnet50 networks and the mutual average teaching method (MUTUAL MEAN-TEACHING, MMT) for pedestrian re-identification are used as reference algorithms. The test data sets use Market-1501, DukeMTMC-reid and MSMT data sets respectively. On Duke-to-Market, compared with the reference algorithm MMT, the present invention obtains an improvement of Rank-1 / mAP=+5.8% / 8.5%. As shown in Table 1.

[0068] Table 1

[0069]

[0070]

[0071] Embodiment 2:

[0072] A dual-branch neural network pedestrian re-identification device based on multi-granularity feature fusion, comprising:

[0073] Data acquisition unit: used to obtain source domain and target domain data sets for person re-identification;

[0074] Construction unit: used to construct a parallel training architecture, which includes four feature extraction networks with the same network structure, which are used to generate hard labels and soft labels respectively. Each feature extraction network includes two branches: a global branch and a feature fusion branch, and the source domain data is input into the parallel training architecture;

[0075] The first processing unit: used for the global branch to use ResNet-50 to learn the global feature c5 of each image in the source domain data. The step size of res_conv5x in ResNet-50 is 1, and the output size is [16,8];

[0076] The second processing unit is used to fuse the c2, c3, c4, and c5 features in the global branch ResNet-50 using a pyramid network, where c2 is the output result after the input passes through res2, with a size of [64, 32] and 256 channels, c3 is the result after c2 passes through res3, with a size of [32, 16] and 512 channels, c4 is the result after c3 passes through res4, with a size of [16, 8] and 1024 channels, and c5 is the result after c4 passes through res4. The result after res5 is of size [16,8], 2048 channels, and the output is p2, p3, p4. P4 is obtained by adding c4 after upsampling and adjusting the channel of c5, with a size of [16,8], 256 channels. P3 is obtained by adding c3 after upsampling and adjusting the channel of p4, with a size of [32,16], 256 channels. P2 is obtained by adding c2 after upsampling and adjusting the channel of p3, with a size of [64,32], 256 channels.

[0077] The third processing unit is used to adjust the size of p2, p3, and p4 to [16,8] by downsampling to obtain p2', p3', and p4';

[0078] The fourth processing unit is used to splice p2', p3', and p4' along the channel, compress the channel to 1024, and obtain c4_f;

[0079] The fifth processing unit is used to input c4_f into the res_conv5x module to obtain c5_f. The parameters of the res_conv5x module are the same as those of the res_conv5x module in the global branch.

[0080] The sixth processing unit is used to constrain the entire network structure using a loss function;

[0081] The seventh processing unit is used to determine whether the number of iterations set during network training has been reached. If yes, the eighth processing unit is executed; if no, the first processing unit is executed;

[0082] The eighth processing unit: fuses c5 and c5_f obtained from each pedestrian image to obtain the feature feature_r used for pedestrian re-identification;

[0083] Result output unit: used to identify the pedestrian using a distance metric for each pedestrian's feature feature_r and output the recognition result.

[0084] Embodiment three:

[0085] An electronic device comprising

[0086] at least one processor, and

[0087] at least one memory in communication with the processor, wherein:

[0088] The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the above-mentioned dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion.

[0089] Embodiment 4:

[0090] A non-volatile computer-readable storage medium, when the computer-executable instructions are executed by one or more processors, enables the processors to execute the above-mentioned dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A dual-branch neural network person re-identification method based on multi-granularity feature fusion, characterized by: The steps include: Step 1: Obtain source and target domain datasets for person re-identification; Step 2: Construct a parallel training architecture, which includes four feature extraction networks with the same network structure, which are used to generate hard labels and soft labels respectively. Each feature extraction network includes two branches: a global branch and a feature fusion branch. The source domain data is input into the parallel training architecture. Step 3: The global branch uses ResNet-50 to learn the global feature c5 of each image in the source domain data. The step size of res_conv5x in ResNet-50 is 1, and the output size is [16, 8]; Step 4: The fusion branch uses a pyramid network to fuse the c2, c3, c4, and c5 features in the global branch ResNet-50, where c2 is the output result after the input passes through res2, with a size of [64, 32] and 256 channels, c3 is the result after c2 passes through res3, with a size of [32, 16] and 512 channels, c4 is the result after c3 passes through res4, with a size of [16, 8] and 1024 channels, and c5 is the result after c4 passes through res The result after s5 is of size [16,8], 2048 channels, and the outputs are p2, p3, and p4. P4 is obtained by adding c4 after upsampling and adjusting the channels of c5, with a size of [16,8], 256 channels. P3 is obtained by adding c3 after upsampling and adjusting the channels of p4, with a size of [32,16], 256 channels. P2 is obtained by adding c2 after upsampling and adjusting the channels of p3, with a size of [64,32], 256 channels. Step 5: Resize p2, p3, and p4 to [16,8] by downsampling to obtain p2', p3', and p4'; Step 6: Splice p2', p3', and p4' along the channel, compress the channel to 1024, and obtain c4_f; Step 7: Input c4_f into the res_conv5x module to obtain c5_f. The parameters of the res_conv5x module are the same as those in the global branch. Step 8: Use the loss function to constrain the entire network structure; Step 9: Determine whether the number of iterations set during network training has been reached. If yes, execute step 10; if no, execute step 3; Step 10: Fuse c5 and c5_f obtained from each pedestrian image to obtain the final feature feature_r for pedestrian re-identification; Step 11: Use the distance measurement method to determine the identity of each pedestrian's feature feature_r and output the recognition result.

2. The method for pedestrian re-identification based on dual-branch neural network with multi-granularity feature fusion according to claim 1 is characterized in that: In step 8, the loss function is expressed by the following expression: is the hard label identity loss in the feature extraction network, is the hard label triplet loss in the feature extraction network, is the soft label identity loss in the feature extraction network, is the soft label triplet loss in the feature extraction network, and is the weight coefficient, is the classification of the target domain under the four networks, They are the feature output of the main branch and the feature output of the fusion branch under the same network. They represent the current feature, the most difficult positive sample, and the most difficult negative sample respectively. The most difficult positive sample represents the feature that is the most dissimilar to the current one in the current training batch data. The most difficult negative sample represents the feature that is the most similar to the current one in the current training batch data. express is the true weight matrix of the kth person, E (T) is the balance coefficient between the Tth iteration and the T-1th iteration of each model weight parameter.

3. The method for pedestrian re-identification based on dual-branch neural network with multi-granularity feature fusion according to claim 1 is characterized in that: In step 10, the fusion method is to take the maximum value of each dimension of c5 and c5_f as the final feature_r.

4. A dual-branch neural network pedestrian re-identification device based on multi-granularity feature fusion, characterized by: include: Data acquisition unit: used to obtain source domain and target domain data sets for person re-identification; Construction unit: used to construct a parallel training architecture, which includes four feature extraction networks with the same network structure, which are used to generate hard labels and soft labels respectively. Each feature extraction network includes two branches: a global branch and a feature fusion branch, and the source domain data is input into the parallel training architecture; The first processing unit: used for the global branch to use ResNet-50 to learn the global feature c5 of each image in the source domain data. The step size of res_conv5x in ResNet-50 is 1, and the output size is [16,8]; The second processing unit is used to fuse the c2, c3, c4, and c5 features in the global branch ResNet-50 using a pyramid network, where c2 is the output result after the input passes through res2, with a size of [64, 32] and 256 channels, c3 is the result after c2 passes through res3, with a size of [32, 16] and 512 channels, c4 is the result after c3 passes through res4, with a size of [16, 8] and 1024 channels, and c5 is the result after c4 passes through res4. The result after res5 is of size [16,8], 2048 channels, and the output is p2, p3, p4. P4 is obtained by adding c4 after upsampling and adjusting the channel of c5, with a size of [16,8], 256 channels. P3 is obtained by adding c3 after upsampling and adjusting the channel of p4, with a size of [32,16], 256 channels. P2 is obtained by adding c2 after upsampling and adjusting the channel of p3, with a size of [64,32], 256 channels. The third processing unit is used to adjust the size of p2, p3, and p4 to [16,8] by downsampling to obtain p2', p3', and p4'; The fourth processing unit is used to splice p2', p3', and p4' along the channel, and then compress the channel to 1024 to obtain c4_f; The fifth processing unit is used to input c4_f into the res_conv5x module to obtain c5_f. The parameters of the res_conv5x module are the same as those of the res_conv5x module in the global branch. The sixth processing unit is used to constrain the entire network structure using a loss function; The seventh processing unit is used to determine whether the number of iterations set during network training has been reached. If yes, the eighth processing unit is executed; if no, the first processing unit is executed; The eighth processing unit: fuses c5 and c5_f obtained from each pedestrian image to obtain the feature feature_r used for pedestrian re-identification; Result output unit: used to identify the pedestrian using a distance metric for each pedestrian's feature feature_r and output the recognition result.

5. An electronic device, characterized in that: include at least one processor, and at least one memory in communication with the processor, wherein: The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion as described in any one of claims 1-3.

6. A non-volatile computer-readable storage medium, when the computer executable instructions are executed by one or more processors, enables the processor to execute the dual-branch neural network pedestrian re-identification method based on multi-granularity feature fusion described in any one of claims 1-3.

Images