Footprint recognition system based on cross-component metrics and multi-scale adaptive fusion

Abstract

The application provides a footprint recognition system based on cross-component measurement and multi-scale adaptive fusion, relates to the technical field of footprint recognition, and comprises a preprocessing module, a main network module, a feature segmentation module and a feature fusion module.The preprocessing module is used for performing data enhancement and size standardization processing on a footprint image to obtain initial footprint features.The main network module is used for performing parallel feature extraction on the initial footprint features to obtain auxiliary features and multi-level backbone features, reducing and projecting high-dimensional backbone features to a fusion space, and performing fusion processing based on an attention mechanism to obtain first fusion features.The feature segmentation module is used for performing multi-granularity segmentation processing on the first fusion features, backbone features outside the highest level and auxiliary features respectively to obtain multiple branch features.The feature fusion module is used for fusing the branch features to obtain second fusion features.The footprint comparison module is used for comparing footprints based on the second fusion features to determine the person to whom the footprint belongs.Through the above footprint recognition system, accurate recognition of footprints can be realized.

Description

Technical Field

[0001] This application relates to the field of footprint recognition technology, and in particular to a footprint recognition system based on cross-component measurement and multi-scale adaptive fusion. Background Technology

[0002] Currently, traditional footprint recognition technology faces two major challenges: first, the limitation of information sources, meaning that a single image cannot provide sufficiently comprehensive feature data; and second, strong interference from environmental factors, where any changes in lighting, viewing angle, and scale can lead to image feature distortion or instability. These two problems work together to directly weaken the reliability of footprint recognition results and easily lead to erroneous judgments.

[0003] In recent years, deep learning-based footprint recognition methods have significantly improved accuracy and robustness, becoming one of the mainstream approaches. However, most deep learning methods tend to use the entire footprint image as input and employ a single backbone network for global feature extraction. While this approach comprehensively considers image information, it often fails to effectively extract local shoe print texture information, such as highly discriminative subtle features like arch curvature, toe shape, and pressure distribution, when faced with complex backgrounds, resulting in poor footprint recognition accuracy. Summary of the Invention

[0004] In view of this, this application proposes a footprint recognition system based on cross-component measurement and multi-scale adaptive fusion.

[0005] In a first aspect, this application provides a footprint recognition system based on cross-component measurement and multi-scale adaptive fusion, the footprint recognition system comprising:

[0006] The preprocessing module is used to perform data augmentation and size normalization on the footprint images to be identified, so as to obtain initial footprint features;

[0007] The main network module is used to extract features from the initial footprint features through a dual backbone network, obtaining multiple backbone features at different abstraction levels and auxiliary features with a predetermined number of channels. The backbone features with a number of channels not less than the fusion dimension are dimensionality-reduced and projected into the fusion space, and fusion processing is performed based on an attention mechanism to obtain a first fused feature corresponding to the highest-level backbone feature. The dimension of the fusion space is the fusion dimension, and the number of channels of the auxiliary features is different from the number of channels of each backbone feature.

[0008] The feature segmentation module is used to perform multi-granularity adaptive segmentation processing on the first fused feature, the main features at other scales except for the highest-level main feature, and the auxiliary features to obtain multiple branch features;

[0009] The feature fusion module is used to fuse the features of each branch to obtain a second fused feature;

[0010] The footprint comparison module is used to compare the second fused feature with the features in the feature database to determine the person to whom the footprint belongs and the similarity.

[0011] In one embodiment, the preprocessing module includes: a normalization submodule and a data augmentation submodule;

[0012] The standardization submodule is used to convert the footprint image to be identified into a three-channel color image, maintain the three-channel information of the color image, adjust all images to a preset standard resolution, and normalize the image pixel values ​​using the standard mean and standard deviation of the ImageNet pre-trained model.

[0013] The data augmentation submodule is used to perform data augmentation processing on the image output by the standardization submodule to obtain the initial footprint features.

[0014] In one embodiment, the main network module includes: a residual network branch, an efficient network branch, and a cross-scale residual attention fusion submodule;

[0015] The residual network branch is used to process the initial footprint features through a hierarchical coding mechanism to extract backbone features at multiple abstract levels;

[0016] The efficient network branch is used to process the initial footprint features to obtain the auxiliary features;

[0017] The cross-scale residual attention fusion submodule is used to project the backbone features with a channel number not less than the fusion dimension into the fusion space by dimensionality reduction. It uses the highest-level backbone feature as the query and backbone features of other scales as keys and values ​​to perform cross-scale attention calculation in the fusion space. The calculated residual information is then upsampled to restore the highest channel number to obtain the restored residual information. The restored residual information is then added element-wise to the original backbone feature of the highest level and subjected to normalization and activation processing to obtain the first fusion feature.

[0018] In one embodiment, the residual network branch is a ResNet50 network branch;

[0019] The ResNet50 network branch is used to extract features from the initial footprint features through five encoder stages, resulting in backbone features with 64, 256, 512, 1024, and 2048 channels, respectively.

[0020] In one embodiment, the efficient network branch is an EfficientNetV2-S network branch;

[0021] The EfficientNetV2-S network branch is used to extract features from the initial footprint features to obtain 128-channel auxiliary features.

[0022] In one embodiment, the feature segmentation module includes: a segmentation submodule and a post-processing submodule.

[0023] The segmentation submodule is used to segment the first fused feature, the main features at other scales except the highest-level main feature, and the auxiliary features using a dynamic multi-segmentation strategy to obtain multiple segmentation features. The segmentation submodule dynamically predicts the optimal segmentation boundary through a learnable neural network.

[0024] The post-processing submodule includes an RSBN (Randomly Discarded Scale Branches Batch Normalization) unit, a pooling unit, and an output unit. The RSBN unit is used to extract multi-scale features for each segmentation feature and perform randomly discarded scale branches batch normalization on the extracted features. The pooling unit is used to perform feature aggregation on the output features of the RSBN unit using a generalized mean pooling strategy. The output unit is used to perform two-dimensional batch normalization and flattening operations on the output features of the pooling unit and output the features of each branch. The generalized mean pooling strategy uses learnable power parameters.

[0025] In one embodiment, the feature fusion module includes: a splicing submodule, a dimensionality reduction submodule, and a classification submodule;

[0026] The splicing submodule is used to splice the branch features according to the channel dimension to obtain spliced ​​features;

[0027] The dimensionality reduction submodule is used to reduce the dimensionality of the spliced ​​features through a fully connected layer to obtain dimensionality-reduced global features;

[0028] The classification submodule is used to perform one-dimensional batch normalization on the dimensionality-reduced global features, and flatten the feature dimensions through a flattening operation, perform a linear transformation on the flattened features, and output the second fused features.

[0029] In one embodiment, the footprint comparison module is further configured to calculate the similarity between the second fused feature and features in the feature database using Euclidean distance or cosine similarity, and determine the person to whom the footprint belongs based on the calculated similarity.

[0030] Secondly, this application also provides a footprint recognition method, the footprint recognition method comprising:

[0031] Obtain a training sample set and a footprint recognition system as described in the first aspect;

[0032] The footprint recognition system is trained based on the training sample set and the target loss function to obtain a target footprint comparison model, wherein the target loss function is a weighted summation function of the comparison center marginal loss function and the label smoothing cross-entropy loss.

[0033] The footprint image to be identified is input into the target footprint comparison model to obtain the person to whom the footprint belongs and the similarity.

[0034] Thirdly, this application also provides a footprint recognition method based on cross-component measurement and multi-scale adaptive fusion, including:

[0035] The footprint images to be identified are subjected to data augmentation and size normalization to obtain initial footprint features;

[0036] The initial footprint features are extracted using a dual-backbone network to obtain multiple backbone features at different abstraction levels and auxiliary features with a predetermined number of channels. The backbone features with a number of channels not less than the fusion dimension are dimensionality-reduced and projected into the fusion space, and fusion processing is performed based on an attention mechanism to obtain the first fused feature corresponding to the highest-level backbone feature. The dimension of the fusion space is the fusion dimension, and the number of channels of the auxiliary features is different from the number of channels of each backbone feature.

[0037] The first fused feature, the main features at other scales excluding the highest-level main feature, and the auxiliary features are subjected to multi-granularity adaptive segmentation processing to obtain multiple branch features;

[0038] The branch features are fused together to obtain the second fused feature;

[0039] The second fused feature is compared with the features in the feature database to determine the person to whom the footprint belongs and the similarity.

[0040] The footprint recognition system based on cross-component measurement and multi-scale adaptive fusion proposed in this application has the following advantages over related technologies:

[0041] 1. The footprint recognition system of this application first performs data augmentation and size standardization on the footprint images through a preprocessing module. This not only expands the sample diversity to improve the model's generalization ability, but also unifies the input specifications to eliminate interference from size differences, providing high-quality, standardized initial footprint features for subsequent feature extraction. Then, the dual backbone network of the main network module extracts backbone features at different abstraction levels and auxiliary features with different channel numbers in parallel. This breaks the limitation of single backbone network feature extraction with a single dimension, covering full-dimensional information from low-level texture to high-level semantics. Furthermore, dimensionality reduction projection significantly reduces computational complexity, and the attention mechanism automatically allocates feature weights, strengthening highly discriminative local features and suppressing background interference, generating a first fusion feature that combines global expression and detail capture capabilities.

[0042] 2. The feature segmentation module performs multi-granularity adaptive segmentation on the first fusion feature, other scale backbone features, and auxiliary features. This uses large granularity to adapt to the global structure and small granularity to capture local details, avoiding the omission of key information caused by single granularity and fully releasing the representational potential of each feature. Then, the feature fusion module integrates the multi-granularity branch features across components, scales, and dimensions to achieve feature complementarity between the global structure and local details, generating a second fusion feature with a more comprehensive representation. Finally, the footprint comparison module accurately compares this highly recognizable feature with database features, effectively reducing the false positive and false negative rates in complex backgrounds, significantly improving the accuracy and robustness of footprint recognition, and achieving accurate determination of the person to whom the footprint belongs and the similarity. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the structure of a footprint recognition system based on cross-component measurement and multi-scale adaptive fusion in one embodiment of this application;

[0045] Figure 2 This is a schematic diagram of the architecture of a footprint recognition system based on cross-component measurement and multi-scale adaptive fusion in one embodiment of this application;

[0046] Figure 3 This is a schematic diagram of the architecture of a ResNet50 network branch in one embodiment of this application;

[0047] Figure 4 This is a schematic diagram of the architecture of an EfficientNetV2-S network branch in one embodiment of this application;

[0048] Figure 5 This is a schematic diagram of the architecture of the cross-scale residual attention fusion submodule in one embodiment of this application;

[0049] Figure 6 This is a schematic diagram of the architecture of the feature segmentation module in one embodiment of this application;

[0050] Figure 7 This is a schematic diagram of the architecture of the post-processing submodule in one embodiment of this application;

[0051] Figure 8 This is a schematic diagram of the process for calculating the loss using the target loss function in one embodiment of this application;

[0052] Figure 9 This is a flowchart illustrating a footprint recognition method based on cross-component measurement and multi-scale adaptive fusion in one embodiment of this application.

[0053] Figure 10 This is a schematic flowchart of a footprint recognition method in one embodiment of this application.

[0054] Explanation of reference numerals in the attached figures:

[0055] 10- Footprint recognition system based on cross-component measurement and multi-scale adaptive fusion, 11- Preprocessing module, 12- Main network module, 13- Feature segmentation module, 14- Feature fusion module, 15- Footprint comparison module. Detailed Implementation

[0056] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0057] In some embodiments, such as Figure 1 As shown, this application provides a footprint recognition system 10 based on cross-component measurement and multi-scale adaptive fusion. The footprint recognition system 10 includes: a preprocessing module 11, a main network module 12, a feature segmentation module 13, a feature fusion module 14, and a footprint comparison module 15.

[0058] The preprocessing module 11 is used to perform data augmentation and size normalization on the footprint image to be identified, so as to obtain the initial footprint features.

[0059] The data augmentation stage addresses common issues in field-collected footprint images, such as blurriness, uneven lighting, partial defects, and cluttered backgrounds. It employs strategies like random rotation, translation, flipping, brightness and contrast adjustment, and noise addition to effectively expand sample diversity and allow the model to adapt to varying footprint morphology in complex field environments, reducing the risk of overfitting due to limited sample size and scene constraints. Size standardization, on the other hand, eliminates size differences in footprint images caused by different acquisition devices and angles. It uniformly scales all images to a preset fixed size, ensuring consistent feature data dimensions input to the subsequent main network module 12 and preventing feature extraction misalignment and feature dimension confusion caused by size mismatches.

[0060] The main network module 12 is used to extract features from the initial footprint features through a dual backbone network to obtain multiple backbone features at different abstraction levels and auxiliary features with a predetermined number of channels. The backbone features with a number of channels not less than the fusion dimension are dimensionality-reduced and projected into the fusion space, and fusion processing is performed based on the attention mechanism to obtain the first fusion feature corresponding to the highest-level backbone feature. The dimension of the fusion space is the fusion dimension, and the number of channels of the auxiliary features is different from the number of channels of each backbone feature.

[0061] The main network module 12 extracts initial footprint features in parallel using a dual-backbone network, breaking the limitation of traditional single-backbone networks in terms of single-dimensional feature extraction. One backbone network progressively extracts multiple backbone features at different abstract levels, from low to high, while the other backbone network simultaneously extracts auxiliary features with a predetermined number of channels. The number of channels for these auxiliary features is intentionally designed to differ from the number of channels for each backbone feature, supplementing the details not covered by the backbone features from a differentiated perspective, thus forming complementary features to enrich the overall feature representation dimension. For backbone features with a number of channels not less than the fusion dimension, the main network module 12 reduces their dimensionality to a preset fusion space using methods such as linear projection, achieving alignment of multi-scale backbone features in the same dimensional space, thereby significantly reducing computational complexity. Based on this, an attention mechanism is introduced for fusion processing. By dynamically calculating the weight coefficients of each backbone feature and auxiliary feature, the contribution ratio of highly discriminative features such as arch curvature and pressure distribution is automatically enhanced, while suppressing the interference of invalid information from complex field backgrounds. Finally, the first fusion feature corresponding to the highest-level backbone feature is generated.

[0062] The feature segmentation module 13 is used to perform multi-granularity adaptive segmentation on the first fusion feature, other scales of main features (excluding the highest-level main features), and auxiliary features, respectively, to obtain multiple branch features. The feature segmentation module 13 overcomes the limitations of single-granularity segmentation in adapting to different scales and semantic features through multi-granularity adaptive segmentation, fully releasing the representational potential of various features. For the first fusion feature, which combines global semantics and local details, large-granularity segmentation can be used. For other scale main features, medium- or small-granularity segmentation can be used. For auxiliary features with varying channel numbers, small-granularity segmentation can also be used to mine their unique dimensions and supplementary details, forming a complement to the main features and the first fusion feature. The segmentation ratio can be adaptively adjusted according to the feature content.

[0063] The feature fusion module 14 is used to fuse the features of each branch to obtain the second fused feature. By integrating the multi-granularity branch features through the feature fusion module 14, the feature complementarity between global structure and local details is achieved, thereby generating a second fused feature with a more comprehensive representation.

[0064] The footprint comparison module 15 compares the second fused feature with features in the feature database to determine the person to whom the footprint belongs and the similarity score. The footprint comparison module 15 first performs feature alignment processing on the second fused feature and the database features to eliminate feature offset issues caused by different collection scenarios and preprocessing differences. Then, it calculates the similarity between the second fused feature and the features in the feature database using an appropriate metric algorithm (such as cosine similarity or Euclidean distance). Based on the calculated similarity score, the person to whom the footprint belongs is determined. Specifically, a reasonable similarity threshold can be set. When the similarity between the feature to be identified and a registered feature in the database is higher than the threshold, they are determined to belong to the same subject; otherwise, they are classified as mismatched. Finally, a clear determination of the person to whom the footprint belongs and the specific similarity score are output.

[0065] The aforementioned footprint recognition system first performs data augmentation and size standardization on the footprint images through the preprocessing module 11. This expands sample diversity to improve the model's generalization ability and unifies input specifications to eliminate interference from size differences, providing high-quality, standardized initial footprint features for subsequent feature extraction. Then, the dual-backbone network of the main network module 12 extracts backbone features at different abstraction levels and auxiliary features with different channel numbers in parallel. This breaks the limitation of single-backbone network feature extraction with a single dimension, covering full-dimensional information from low-level texture to high-level semantics. Furthermore, dimensionality reduction projection significantly reduces computational complexity, and the attention mechanism automatically allocates feature weights, strengthening highly discriminative local features and suppressing background interference, generating a first fusion feature that combines global expression and detail capture capabilities. The feature segmentation module 13 performs multi-granularity adaptive segmentation on the first fusion feature, other scale backbone features, and auxiliary features. This allows for large-granularity adaptation to the global structure and small-granularity capture of local details, avoiding the omission of key information caused by single-granularity segmentation and fully releasing the representational potential of each feature. The feature fusion module 14 then integrates the multi-granularity branch features across components, scales, and dimensions, achieving feature complementarity between the global structure and local details, and generating a second fusion feature with a more comprehensive representation. Finally, the footprint comparison module 15 accurately compares this highly recognizable feature with database features, effectively reducing the false positive and false negative rates in complex backgrounds, significantly improving the accuracy and robustness of footprint recognition, and achieving accurate determination of the person to whom the footprint belongs and the similarity.

[0066] In some embodiments, the preprocessing module 11 includes a standardization submodule and a data augmentation submodule.

[0067] The standardization submodule is used to convert the footprint images to be identified into three-channel color images, preserve the three-channel information of the color images, uniformly adjust all images to a preset standard resolution, and normalize the image pixel values ​​using the standard mean and standard deviation of the ImageNet pre-trained model.

[0068] The data augmentation submodule is used to perform data augmentation processing on the image output by the normalization submodule to obtain the initial footprint features.

[0069] The preset standard resolution can be 320×160 pixels.

[0070] In the application, the normalization submodule converts the original footprint image into a 3-channel image, preserving the three-channel information and ensuring all inputs have consistent channel specifications. Then, all images are uniformly adjusted to a standard resolution of 320×160 pixels using bicubic interpolation to ensure all input samples have consistent size specifications while maintaining image quality and avoiding detail loss. Finally, the pixel values ​​are normalized using the standard mean [0.485, 0.456, 0.406] and standard deviation [0.229, 0.224, 0.225] of the ImageNet pre-trained model to improve the model's adaptability to different lighting conditions. The normalization formula is:

[0071]

[0072] in, The mean, The standard deviation is 32, 3, 320, 160. The final input feature map shape is [32, 3, 320, 160], where 32 is the batch size, 3 is the RGB three-channel, and 320×160 is the image resolution.

[0073] The data augmentation submodule can apply the following augmentation techniques sequentially during the training phase: random horizontal flip (50% probability), random color transformation (brightness variation of 0.2, contrast variation of 0.15, saturation variation of 0.1, and hue variation of 0.1), random rotation (within the range of -10° to +10°), padding (padding a 10-pixel border around the image), random cropping (cropping a 320×160-pixel region from the padded image), and random erasing (erasing 2% to 50% of the image with a 50% probability) to increase the diversity of the dataset and improve the robustness of the model.

[0074] In some embodiments, the main network module 12 includes: a residual network branch, an efficient network branch, and a cross-scale residual attention fusion submodule.

[0075] The residual network branch is used to process the initial footprint features through a hierarchical coding mechanism to extract backbone features at multiple abstract levels.

[0076] The residual network branch processes the standardized initial footprint features obtained after preprocessing through a hierarchical encoding mechanism, extracting backbone features at multiple abstract levels. Relying on the core architectural advantages of residual connections, this branch effectively solves the gradient vanishing or gradient exploding problem in the training process of deep networks, providing technical support for the stable operation of the hierarchical encoding mechanism.

[0077] Its hierarchical encoding process follows a feature abstraction logic from shallow to deep: the shallow encoding layer focuses on extracting low-level visual features such as edge contours, sole texture direction, and local bumps and depressions in the footprint image through basic convolutional operations. These features retain the most original details of the footprint. The middle encoding layer, based on the shallow features, further mines highly discriminative local features such as arch curvature, toe shape, and toe spacing through residual block stacking operations and feature fusion. These features are the core basis for distinguishing different individual footprints. The deep encoding layer integrates and abstracts the middle and shallow features to generate high-level semantic features that can characterize the overall contour structure of the footprint, the trend of sole pressure distribution, and the overall shape proportions. In the entire hierarchical encoding process, the cross-layer information transfer mechanism of residual connections allows the detailed features of the shallow layer to flow effectively to the middle and deep networks, avoiding the loss of local subtle information caused by the layer-by-layer abstraction of features.

[0078] The efficient network branch processes the initial footprint features to obtain auxiliary features. As an important component of the dual-backbone network, the efficient network branch collaborates in parallel with the residual network branch to perform lightweight and differentiated feature extraction on the standardized initial footprint features output after preprocessing. This results in auxiliary features with a different number of channels than the backbone features at each abstraction level. This branch abandons the deep hierarchical encoding mode of stacking residual network branches, adopting a lightweight network architecture design. This significantly reduces computational complexity and improves feature extraction speed while avoiding feature extraction redundancy with the residual network branch. It focuses on capturing unique discriminative features that are easily masked or missed by the residual network branch during the layer-by-layer abstraction process. These features often lack obvious hierarchical abstraction patterns and are difficult to capture effectively by the hierarchical encoding mechanism of the residual network, yet they are of significant value in distinguishing individual footprints.

[0079] The cross-scale residual attention fusion submodule is used to reduce the dimensionality of the backbone features with a number of channels not less than the fusion dimension and project them into the fusion space. The highest-level backbone features are used as queries, and the backbone features of other scales are used as keys and values. Cross-scale attention calculation is performed in the fusion space. The calculated residual information is upsampled to restore the highest number of channels to obtain the restored residual information. The restored residual information is added element-wise to the original backbone features of the highest level, and then normalized and activated to obtain the first fusion feature.

[0080] It is understandable that the cross-scale residual attention fusion submodule achieves accurate fusion of backbone features at different abstraction levels through a three-step strategy of dimensionality reduction alignment, attention filtering, and residual enhancement. First, for backbone features with a channel number greater than or equal to the fusion dimension, a linear projection operation is used to map them to a preset fusion space. This eliminates the dimensionality heterogeneity problem caused by differences in the number of channels among backbone features at different scales, ensuring that all features participating in the fusion are in the same computable dimension, thus laying the foundation for cross-scale feature interaction. Secondly, using the highest-level backbone features representing the global structure of the footprint as query features, and other-scale backbone features carrying local details such as edge texture and arch curvature as key and value features, cross-scale attention computation is performed. With the dynamic allocation of attention weights, the global features guide the filtering of highly discriminative information in local features, while suppressing invalid interference from complex backgrounds, generating residual information that focuses on core features. Subsequently, an upsampling operation is used to restore the number of channels in the residual information to the same level as the original backbone features of the highest level, solving the feature compression loss problem that may occur during dimensionality reduction. Finally, the restored residual information is added element-wise to the original backbone features of the highest level. The advantages of residual connections are used to preserve the global semantic expression of the highest-level features, while incorporating detailed information from other-scale features. Then, a normalization operation is performed to stabilize the feature distribution, and an activation function is used to introduce nonlinear feature transformation, ultimately outputting a first fusion feature that combines global structural integrity and local detail discriminability.

[0081] In some embodiments, such as Figure 2 and Figure 3 As shown, the residual network branch is a ResNet50 network branch. The ResNet50 network branch is used to extract features from the initial footprint features through five encoder stages, resulting in backbone features with 64, 256, 512, 1024, and 2048 channels, respectively.

[0082] In applications, the ResNet50 network branch can be based on ImageNet pre-trained weights. By modifying the stride of the last convolutional layer in layer 4 from 2 to 1, higher spatial resolution is maintained. Then, multi-level features are extracted through five encoder stages. Stage 0 outputs a feature map F0 of [32, 64, 80, 40] through a 7×7 convolutional layer (64 channels, stride=2, padding=3) and a 3×3 max-pooling layer (stride=2), capturing basic edge and texture information. Stage 1 outputs a feature map F1 of [32, 256, 80, 40] through layer 1 (3 Bottleneck modules, each containing a 1×1→3×3→1×1 convolutional sequence), extracting local structural features. Stage 2 outputs a feature map F2 of [32, 512, 40, 20] through layer 2 (4 Bottleneck modules, downsampled with stride=2), obtaining moderate semantic information. Stage 3 outputs a feature map F3 of [32, 1024, 20, 10] through layer 3 (6 Bottleneck modules, downsampled with stride=2), capturing high-level semantic features. Stage 4 outputs a feature map F4 of [32, 2048, 10, 5] through layer 4 (3 Bottleneck modules, with stride modified to 1 to maintain resolution), obtaining global contextual information.

[0083] In some embodiments, such as Figure 2 and Figure 4 As shown, the efficient network branch is the EfficientNetV2-S network branch. The EfficientNetV2-S network branch is used to extract features from the initial footprint features, obtaining 128-channel auxiliary features. It should be noted that the EfficientNetV2-S network branch in this embodiment only uses the first 5 feature extraction modules of EfficientNetV2-S (including a stem layer, one Fused-MBConv1 block, two Fused-MBConv4 blocks, and one MBConv4 block), rather than using the complete network structure. The stem layer, serving as the initial feature extraction module of the network, includes a 3×3 convolutional layer (stride=2) and a BatchNorm layer, responsible for preliminary downsampling and feature extraction of the input features. The Fused-MBConv module (which combines extended convolution and depthwise separable convolution) accelerates training by fusing operators in the early stages, while the MBConv module (moving inverse residual convolution) utilizes the inverse residual structure to expand the receptive field while maintaining the model's lightweight nature. Both are core components of EfficientNetV2, achieving efficient feature extraction through lightweight convolution operations.

[0084] It is understandable that the EfficientNetV2-S network branch, as a supplementary feature extractor, outputs a feature map F5 of [32,128,20,10] in the first three feature extraction stages through its innovative multi-scale deep convolution and MBConv (moving inverse bottleneck convolution) modules. This provides fine-grained detail information that ResNet50 may ignore, forming a complementary feature representation system with two branches.

[0085] Based on this embodiment and the previous embodiment, such as Figure 5 As shown, the cross-scale residual attention fusion submodule can perform cross-scale fusion of intermediate layer features F1, F2, F3, and F4 through the CSRAF module, realizing intelligent integration of feature information from different levels. Specifically, the 256, 512, 1024, and 2048 channel features are first projected onto a 256-dimensional fusion space using 1×1 convolutions, significantly reducing computational complexity. Then, the highest-level 2048-dimensional feature F4 is used as the query, and the other scale features F1, F2, and F3 are used as the key and value. Cross-scale attention calculation is performed in the low-dimensional space. Bilinear interpolation can be used to align F1, F2, and F3 to the spatial dimensions [10, 5] of F4, and then attention weights are calculated. The learned weights generate fused residual information, which represents useful information supplemented from lower-level features to higher-level features. The attention weight calculation formula is:

[0086]

[0087] in, Attention weights (representing the first) The query position for the first... The level of attention given to each key position, with a value range of [0,1]). For the first Query vectors at each position, For the first A key vector at each position, The dimension of the key vector. The total number of key vectors. For exponential functions (mapping the dot product result to the positive number domain), For summation operations (normalizing all key positions), This is a scaling factor (to prevent the gradient from vanishing due to an excessively large dot product result). This is the dot product operation (calculating the similarity between the query vector and the key vector).

[0088] The calculated residual information is upsampled using a 1×1 convolution to restore it to 2048 channels, yielding the restored residual information. Finally, the restored residual information is element-wise added to the original 2048-dimensional feature F4, and post-processed using BatchNorm and ReLU activation functions to obtain the enhanced global feature F4'. This residual connection strategy ensures training stability, preserving the original feature information even if the attention mechanism fails, while significantly enhancing the global expressive power and detail capture ability of the features through cross-scale information interaction.

[0089] In some embodiments, the feature segmentation module 13 includes a segmentation submodule and a post-processing submodule.

[0090] The segmentation submodule is used to segment the first fused feature, the main features at other scales except the highest level main feature, and the auxiliary features using a dynamic multi-segmentation strategy to obtain multiple segmentation features. The segmentation submodule dynamically predicts the optimal segmentation boundary through a learnable neural network.

[0091] The post-processing submodule includes an RSBN (Randomly Discarded Scale Branches Batch Normalization) unit, a pooling unit, and an output unit. The RSBN unit is used to extract multi-scale features for each segmentation feature and perform batch normalization processing on the extracted features by randomly discarding scale branches. The pooling unit is used to perform feature aggregation processing on the output features of the RSBN unit using a generalized mean pooling strategy. The output unit is used to perform two-dimensional batch normalization and flattening operations on the output features of the pooling unit and output the features of each branch. The generalized mean pooling strategy uses learnable power parameters.

[0092] Understandable, such as Figure 6 As shown, the segmentation submodule performs multi-granularity adaptive segmentation processing on the first fused feature, the main features and auxiliary features at other scales except the highest level main feature (such as the aforementioned F0, F1, F2, F3, F4', F5), which can achieve fine decomposition and representation of the feature map.

[0093] For example, the dynamic multi-segmentation strategy includes a two-segmentation strategy and a three-segmentation strategy, applied to six feature maps, and the segmentation ratio prediction formula is:

[0094]

[0095] in, To divide the proportional vector, For the first The relative proportions of each segmentation feature The total number of segmentation features. The parameters are the learnable weight matrix. These are learnable bias parameters. For the input feature map, For input features The result of performing a global pooling operation.

[0096] Then the segmentation submodule calculates the formula based on the actual segmentation size. Determine the size of each segmentation feature, where, For the first The actual size of each segmentation feature For the first The relative proportions of each segmentation feature The length of the feature map in the segmentation dimension. This is the rounding function. To ensure a function with a minimum size of 1, the bi-segmentation and tri-segmentation ratios are adaptively adjusted based on the feature content, avoiding the rigidity of traditional fixed-ratio segmentation.

[0097] like Figure 7 As shown, the post-processing submodule applies the MultiNormGemPool operation to each segmentation feature. This operation first extracts multi-scale features through the RDBN (Randomly Discarded Scale Branches Batch Normalization) module. The multi-scale feature extraction formula is as follows:

[0098]

[0099] in, This is the result of multi-scale feature fusion. For feature splicing operations, For use Features extracted by convolution kernels For input A random dropout mechanism is applied to dynamically discard some scale branches based on the dropout rate; random batch normalization using RDBN is used, and the formula for random selection during training is:

[0100]

[0101] The formula for using the average value in the evaluation is:

[0102]

[0103] in, The features are normalized. For randomly selected batch normalization operations, The total number of batch normalized layers, For the first Batch normalization layer operations.

[0104] Then, the pooling unit applies GeM pooling to the output features of the RDSBN unit, using learnable power-law parameters. For feature aggregation, the GeM pooling formula is:

[0105]

[0106] in, These are the features after GeM pooling. It is a set of pixels within a local region. For the size of the local region, The pixel values ​​within the region. For learnable exponential parameters, GeM pooling is the reciprocal of the power of 1, compared to traditional average pooling. This allows for better preservation of important feature information. Finally, the output unit outputs a 258-dimensional feature vector through BatchNorm2d (two-dimensional batch normalization) and flattening operations, with each segmentation feature generating an independent feature representation; a total of 42 feature branches are formed (6 feature maps × 7 segmentation features), with each branch outputting 258-dimensional features, thus constructing a complete multi-granularity feature extraction system.

[0107] In some embodiments, the feature fusion module 14 includes: a splicing submodule, a dimensionality reduction submodule, and a classification submodule.

[0108] The splicing submodule is used to splice the features of each branch according to the channel dimension to obtain the spliced ​​features.

[0109] The dimensionality reduction submodule is used to reduce the dimensionality of the spliced ​​features through a fully connected layer to obtain the dimensionality-reduced global features.

[0110] The classification submodule is used to perform one-dimensional batch normalization on the dimensionality-reduced global features, and flatten the feature dimensions through a flattening operation. The flattened features are then linearly transformed to output the second fused feature.

[0111] The splicing submodule fuses all the obtained branch features to construct a unified feature representation system, realizing the transformation from multi-granularity local features to globally consistent features. For example, this process can be as follows: the 258-dimensional features of the 42 branches are spliced ​​along the channel dimension to form a high-dimensional fused feature tensor of 258×42. The feature splicing formula is:

[0112]

[0113] in, This is the fused feature matrix after splicing. For feature splicing operations, For the first The feature vectors of each branch The total number of branches, The dimension of the feature vector for each branch, The process effectively integrates feature information from different levels (F0-F5) and different segmentation regions (two-segmentation and three-segmentation) in a 42-row, 258-column real matrix space, preserving rich local details and global context information.

[0114] Then, the dimensionality reduction submodule performs dimensionality reduction on the high-dimensional features through a fully connected layer (feature linear layer), compressing the 258×42=10836 dimensions to 1045 dimensions. The dimensionality reduction transformation formula is as follows:

[0115]

[0116] in, The global feature vector after dimensionality reduction. The projection weight matrix is... The concatenated feature vectors This is the bias vector. This dimensionality reduction operation not only reduces computational complexity. (in , More importantly, it learns the optimal feature combination method through linear transformation, generating a compact and expressive global context feature representation, i.e., dimensionality-reduced global features.

[0117] Finally, the classification submodule generates the final classification result through a linear classification layer. This layer first uses BatchNorm1d (one-dimensional batch normalization) for normalization to ensure the stability of the feature distribution. Then, it flattens the feature dimensions through a Flatten operation. Finally, it outputs 1045-dimensional classification features through a linear transformation. This feature vector contains discriminative information for identity recognition while maintaining sufficient representational power for subsequent retrieval and matching tasks. This feature fusion strategy, through an end-to-end learning approach, automatically discovers the optimal fusion weights between features of different granularities, significantly improving the overall feature expressiveness and discriminative performance.

[0118] In some embodiments, such as Figure 8 As shown, this application also provides a footprint recognition method, which includes the following steps S801 to S803.

[0119] S801: Acquire training sample set and footprint recognition system of any of the above schemes.

[0120] The footprint recognition system can be referred to in the description of the foregoing embodiments, and will not be repeated here.

[0121] S802: Train the footprint recognition system based on the training sample set and the target loss function to obtain the target footprint comparison model. The target loss function is a weighted sum of the comparison center marginal loss function and the label smoothing cross-entropy loss.

[0122] Among them, such as Figure 9 As shown, the calculation process of the CCMLoss (contrast center marginal loss) function can be as follows: First, all input features in the batch are subjected to L2 normalization, mapping the feature vectors to a unit hypersphere. The L2 normalization formula is:

[0123]

[0124] in, For the first The normalized feature vector of each sample For the first The original feature vector of each sample, For the first The Euclidean norm of each sample feature vector is used to ensure numerical stability and the rationality of distance calculation. Then, an orthogonally initialized class embedding matrix is ​​used to store the center vectors of 1045 classes. Orthogonal initialization ensures... That is, the maximum difference between different category centers. The category center extraction formula is:

[0125]

[0126] in, For the first The center vector of each sample corresponding to its class For the category embedding matrix, For the first The category label of each sample, This is the identity matrix. Next, calculate the pairwise distance matrix between all samples within the batch. The formula for calculating the distance matrix is:

[0127]

[0128] in, The distance matrix is... For the first The sample and the first Euclidean distance between samples These are the normalized eigenvectors.

[0129] The distances between similar samples are filtered out using a different-class mask, retaining only the distances between samples from different classes. An infinite value is set in the distance matrix to exclude similar samples. For each sample, the minimum distance to all samples from different classes is extracted as the minimum negative sample distance. Simultaneously, the distance between each sample (anchor point) and its corresponding class center is calculated. The marginal loss is calculated using the ReLU (Modified Linear Unit) activation function. The loss formula for a single sample is:

[0130]

[0131] in, For the first Marginal loss per sample, For marginal parameters, For the first The distance between each sample and the nearest negative sample. For the first The distance between each sample and its class center, and the total loss is:

[0132]

[0133] in, For the total loss of CCMLoss, As the averaging factor, This is a summation over all samples within a batch. The design philosophy of this loss function is: when the distance between the anchor point and the nearest negative sample minus the distance between the anchor point and the class center is less than a marginal value, a loss is incurred, thereby driving the network to learn to move similar samples closer to the class center and dissimilar samples further away from the class center.

[0134] In loss function combination strategies, CCMLoss (weights) Cross-entropy loss (weights) is the primary metric for learning loss. (Provides classification supervision; the formula for label smoothing cross-entropy loss is:)

[0135]

[0136] in, For label smoothing cross-entropy loss, The label smoothing coefficient. For indicator functions, For the first Each sample belongs to category The predicted probability, The total number of categories, Let be the natural logarithm function. The total loss function is: The footprint retrieval and matching stage efficiently compares query features with footprint features in the database, calculates the similarity between features using Euclidean distance or cosine similarity, and comprehensively evaluates retrieval performance by calculating Rank-1, Rank-5, Rank-10 accuracy and mAP (mean average precision). It supports both same-domain retrieval (query and database come from the same dataset) and cross-domain retrieval (query and database come from different datasets), meeting diverse needs in practical applications.

[0137] S803: Input the footprint image to be identified into the target footprint comparison model to obtain the person to whom the footprint belongs and the similarity.

[0138] The aforementioned footprint recognition method first obtains a training sample set and a footprint recognition system, providing a high-quality feature input foundation and a high-performance feature processing framework for model training. Then, it employs a target loss function that is a weighted sum of the contrast center marginal loss function and the label smooth cross-entropy loss. The contrast center marginal loss function effectively improves the intra-class compactness and inter-class separability of footprint features, allowing the model to more accurately capture the differentiated features of different individuals' footprints. The label smooth cross-entropy loss reduces the model's over-reliance on hard labels, alleviating overfitting. The weighted combination of these two functions achieves complementary optimization of the training objectives, enabling the model to focus on core discriminative features while possessing stronger generalization ability during the learning process. Finally, when the footprint image to be recognized is input into the optimized target footprint comparison model, the model can fully leverage the feature processing advantages of the modular architecture and the accurate discrimination capability brought by the loss function training, efficiently completing feature extraction, fusion, and database comparison. This not only effectively resists interference from complex scene backgrounds and partially incomplete footprints, but also accurately outputs the person to whom the footprint belongs and the corresponding similarity, significantly improving the accuracy of footprint recognition.

[0139] In some embodiments, such as Figure 10 As shown, this application also provides a footprint recognition method based on cross-component measurement and multi-scale adaptive fusion, which includes the following steps S1001 to S1005.

[0140] S1001: Perform data augmentation and size normalization on the footprint image to be identified to obtain initial footprint features.

[0141] S1002: Feature extraction is performed on the initial footprint features using a dual-backbone network to obtain multiple backbone features at different abstraction levels and auxiliary features with a predetermined number of channels. The backbone features with a channel number not less than the fusion dimension are dimensionality-reduced and projected into the fusion space, and fusion processing is performed based on an attention mechanism to obtain the first fused feature corresponding to the highest-level backbone feature. Here, the dimension of the fusion space is the fusion dimension, and the number of channels in the auxiliary features differs from the number of channels in each backbone feature.

[0142] S1003: Perform multi-granularity adaptive segmentation processing on the first fusion feature, the main features and auxiliary features at other scales except the highest level main feature, to obtain multiple branch features.

[0143] S1004: The features of each branch are fused to obtain the second fused feature.

[0144] S1005: Compare the second fused feature with the features in the feature database to determine the person to whom the footprint belongs and the similarity.

[0145] It should be noted that the footprint recognition method based on cross-component measurement and multi-scale adaptive fusion provided in this application embodiment and the footprint recognition system based on cross-component measurement and multi-scale adaptive fusion provided in this application embodiment are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned footprint recognition system based on cross-component measurement and multi-scale adaptive fusion, and the repeated parts will not be described again.

[0146] In some embodiments, an electronic device provided in this application includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the footprint recognition method of any of the above schemes.

[0147] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0148] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0149] This application also provides a non-transitory computer storage medium storing a computer program thereon, which, when executed by a processor, implements the footprint recognition method of any of the above-described schemes. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.

[0150] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0151] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A footprint recognition system based on cross-component measurement and multi-scale adaptive fusion, characterized in that, The footprint recognition system includes: The preprocessing module is used to perform data augmentation and size normalization on the footprint images to be identified, so as to obtain initial footprint features; The main network module includes a residual network branch, an efficient network branch, and a cross-scale residual attention fusion submodule. The residual network branch processes the initial footprint features using a hierarchical encoding mechanism to extract backbone features at multiple abstract levels. The efficient network branch processes the initial footprint features to obtain auxiliary features. The cross-scale residual attention fusion submodule projects backbone features with a channel number not less than the fusion dimension into the fusion space, using the highest-level backbone feature as the query and backbone features at other scales as keys and values. Cross-scale attention computation is performed in the fusion space, and the calculated residual information is upsampled to restore the highest channel number to obtain restored residual information. The restored residual information is then added element-wise to the original backbone features at the highest level, and normalized and activated to obtain the first fusion feature. The dimension of the fusion space is the fusion dimension, and the number of channels in the auxiliary features differs from the number of channels in each backbone feature. The residual network branch is a ResNet50 network branch, and the efficient network branch is an EfficientNetV2-S network branch. The feature segmentation module includes a segmentation submodule and a post-processing submodule. The segmentation submodule is used to segment the first fused feature, the backbone features at other scales except the highest-level backbone feature, and the auxiliary features using a dynamic multi-segmentation strategy to obtain multiple segmentation features. The segmentation submodule dynamically predicts the optimal segmentation boundary using a learnable neural network. The post-processing submodule includes an RSBN unit, a pooling unit, and an output unit. The RSBN unit is used to extract features at multiple scales for each segmentation feature and perform batch normalization processing on the extracted features, randomly discarding scale branches. The pooling unit is used to perform feature aggregation processing on the output features of the RSBN unit using a generalized mean pooling strategy. The output unit is used to perform two-dimensional batch normalization and flattening operations on the output features of the pooling unit, outputting multiple branch features. The generalized mean pooling strategy uses learnable power parameters. The feature fusion module is used to fuse the features of each branch to obtain a second fused feature; The footprint comparison module is used to compare the second fused feature with the features in the feature database to determine the person to whom the footprint belongs and the similarity.

2. The footprint recognition system based on cross-component measurement and multi-scale adaptive fusion as described in claim 1, characterized in that, The preprocessing module includes: a standardization submodule and a data augmentation submodule; The standardization submodule is used to convert the footprint image to be identified into a three-channel color image, maintain the three-channel information of the color image, adjust all images to a preset standard resolution, and normalize the image pixel values ​​using the standard mean and standard deviation of the ImageNet pre-trained model. The data augmentation submodule is used to perform data augmentation processing on the image output by the standardization submodule to obtain the initial footprint features.

3. The footprint recognition system based on cross-component measurement and multi-scale adaptive fusion as described in claim 1, characterized in that, The ResNet50 network branch is used to extract features from the initial footprint features through five encoder stages, resulting in backbone features with 64, 256, 512, 1024, and 2048 channels, respectively.

4. The footprint recognition system based on cross-component measurement and multi-scale adaptive fusion as described in claim 3, characterized in that, The EfficientNetV2-S network branch is used to extract features from the initial footprint features to obtain 128-channel auxiliary features.

5. The footprint recognition system based on cross-component measurement and multi-scale adaptive fusion as described in claim 1, characterized in that, The feature fusion module includes: a splicing submodule, a dimensionality reduction submodule, and a classification submodule; The splicing submodule is used to splice the branch features according to the channel dimension to obtain spliced ​​features; The dimensionality reduction submodule is used to reduce the dimensionality of the spliced ​​features through a fully connected layer to obtain dimensionality-reduced global features; The classification submodule is used to perform one-dimensional batch normalization on the dimensionality-reduced global features, and flatten the feature dimensions through a flattening operation, perform a linear transformation on the flattened features, and output the second fused features.

6. The footprint recognition system based on cross-component measurement and multi-scale adaptive fusion as described in claim 1, characterized in that, The footprint comparison module is also used to calculate the similarity between the second fused feature and the features in the feature database using Euclidean distance or cosine similarity, and to determine the person to whom the footprint belongs based on the calculated similarity.

7. A footprint recognition method, characterized in that, The footprint recognition method includes: Obtain a training sample set and the footprint recognition system as described in any one of claims 1 to 6; The footprint recognition system is trained based on the training sample set and the target loss function to obtain a target footprint comparison model, wherein the target loss function is a weighted summation function of the comparison center marginal loss function and the label smoothing cross-entropy loss. The footprint image to be identified is input into the target footprint comparison model to obtain the person to whom the footprint belongs and the similarity.

8. A footprint recognition method based on cross-component measurement and multi-scale adaptive fusion, characterized in that, include: The footprint images to be identified are subjected to data augmentation and size normalization to obtain initial footprint features; The initial footprint features are processed through a hierarchical coding mechanism to extract core features at multiple abstract levels; The initial footprint features are processed to obtain auxiliary features; The backbone features with a channel number not less than the fusion dimension are dimensionally reduced and projected into the fusion space. The highest-level backbone features are used as queries, and the backbone features of other scales are used as keys and values. Cross-scale attention calculation is performed in the fusion space. The calculated residual information is then upsampled to restore the highest channel number to obtain the restored residual information. The recovered residual information is added element-wise to the original backbone features of the highest level, and then normalized and activated to obtain the first fused feature; wherein, the dimension of the fusion space is the fusion dimension, and the number of channels of the auxiliary features is different from the number of channels of each backbone feature; A dynamic multi-segmentation strategy is used to segment the first fused feature, the main features at other scales except the highest-level main feature, and the auxiliary features, respectively, to obtain multiple segmentation features. Multi-scale feature extraction is performed on each segmentation feature, and batch normalization processing with random discarding of scale branches is applied to the extracted features. A generalized mean pooling strategy is used to perform feature aggregation processing on the output features after batch normalization, and two-dimensional batch normalization and flattening operations are performed to output multiple branch features. The optimal segmentation boundary is dynamically predicted through a learnable neural network. The branch features are fused together to obtain the second fused feature; The second fused feature is compared with the features in the feature database to determine the person to whom the footprint belongs and the similarity.

Images