Embedded palm vein image recognition method and system based on YOLO lightweight model

Abstract

The application discloses an embedded palm vein image recognition method and system based on a YOLO light model, which comprises the following steps: first, analyzing the contribution degree of a convolution layer, simplifying a network and replacing a convolution mode to obtain a first network structure; then, removing unimportant connections through a pruning technique to obtain a second structure; extracting features from a teacher network and adjusting parameters by using knowledge distillation to obtain a third structure; adopting a mixed precision quantization strategy to obtain a fourth structure; adding a feature pyramid network to capture multi-scale features to obtain a fifth structure; introducing an anatomy-guided attention mechanism to obtain a sixth structure; designing a loss function to adjust weights to obtain a seventh structure; processing low-contrast images through an adaptive feature enhancement module to finally obtain an eighth network structure; and performing palm vein image recognition through the eighth network structure to obtain a recognition result. The application ensures that features meet the recognition accuracy requirement and realizes efficient and accurate palm vein image recognition.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biological image recognition, and particularly relates to an embedded palm vein image recognition method and system based on a YOLO lightweight model. BACKGROUND

[0002] Palm vein recognition technology has important applications in the field of biometric recognition, but it faces the contradiction between model complexity and recognition accuracy. How to achieve high-precision and low-delay recognition on resource-constrained embedded devices is still a major challenge. Although the traditional YOLO model performs well in target detection, its direct application in palm vein recognition will result in a model that is too large and cannot meet the operation capacity and storage limitations of embedded devices. At the same time, palm vein images usually have low contrast and contain multi-level structures such as main blood vessels, secondary branches, and capillaries. How to effectively extract these features and maintain recognition accuracy is a difficult problem. In addition, the change in vein depth also affects the recognition effect. How to introduce anatomical knowledge into the model to enhance the attention to veins of different depths, balance the contradiction between model size, computational complexity, and recognition accuracy during model optimization, and design a suitable loss function to guide the model to learn hierarchical vein feature representation are all key issues that need to be further studied. Finally, how to deploy the optimized model to embedded devices and achieve real-time recognition while ensuring recognition accuracy is the core challenge faced by existing technical solutions. SUMMARY

[0003] To solve the above technical problems, the application provides an embedded palm vein image recognition method and system based on a YOLO lightweight model. The embedded palm vein image recognition method based on a YOLO lightweight model comprises:

[0004] The network structure of the original YOLO model is obtained, the contribution of each convolutional layer to palm vein recognition is analyzed, the number of redundant layers and filters is determined, depth separable convolution is used to replace standard convolution, and a simplified first network structure is obtained;

[0005] Based on the first network structure, the importance of each connection to vein texture feature extraction is calculated using pruning technology. If the importance is lower than a preset threshold, the corresponding connection is removed, and a pruned second network structure is obtained;

[0006] The feature representation of the palm vein image is extracted from a pre-established teacher network, the feature is transferred to the second network structure through a knowledge distillation method, the student network parameters are adjusted, and a third network structure with preliminary recognition ability is obtained;

[0007] Analyze the sensitivity of each layer in the third network structure to low-contrast palm vein images, adopt a mixed precision quantization strategy, assign a 16-bit wide representation to the sensitive layer and an 8-bit wide representation to the non-sensitive layer, and obtain a quantized fourth network structure;

[0008] By designing a multi-level feature extraction module and adding a feature pyramid network to the fourth network structure, the features of the main blood vessels, secondary branches and capillary networks at different resolutions are captured, thereby obtaining a fifth network structure containing multi-scale information;

[0009] Based on the vein layer features in the fifth network structure, an anatomically guided attention mechanism is introduced. By calculating the correlation between the feature maps of each layer and the vein depth, the model's attention to veins of different depths is obtained, resulting in an optimized sixth network structure.

[0010] A loss function is designed based on the output features of the sixth network structure, and the network weights are adjusted by measuring the hierarchical representation errors of the main blood vessels, secondary branches, and capillary features to obtain a seventh network structure that accurately extracts hierarchical features;

[0011] Obtaining running data of the seventh network structure on the embedded device, amplifying texture details of the low-contrast palm vein image using an adaptive feature enhancement module, determining whether the enhanced features meet recognition accuracy requirements, and obtaining a final eighth network structure; performing palm vein image recognition based on the eighth network structure to obtain an image recognition result.

[0012] Preferably, the process of obtaining the simplified first network structure includes:

[0013] Obtain the network structure of the original YOLO model, decompose the convolutional layers to obtain the parameters and output features of each layer, and determine the preliminary analysis results;

[0014] By calculating the contribution of each convolutional layer output feature to palm vein recognition, redundant layers and unnecessary filters are determined to obtain a redundant judgment set;

[0015] Analyzing the filter usage rate of the redundant decision set using a statistical method, determining the number of redundant layers and filters to be removed, and obtaining a streamlined layer configuration;

[0016] Extracting the retained convolutional layers from the streamlined layer configuration, replacing the standard convolution with depthwise separable convolution, to obtain a converted first network structure;

[0017] Running the palm vein dataset through the first network structure, obtaining the recognition rate and computational complexity, determining whether the preset threshold is met, and obtaining a performance evaluation result;

[0018] If the performance evaluation result is lower than the preset threshold value, the depth separable convolution parameter is adjusted to obtain an optimized network structure, otherwise the first network structure is directly used as the second network structure.

[0019] Preferably, the process of obtaining the second network structure after pruning comprises:

[0020] The weight data of each connection in the first network structure is obtained through a pruning technique, the contribution of each connection to the vein texture is calculated by using a convolutional neural network, and an importance value is obtained.

[0021] The importance value is compared with a preset threshold value, if the importance value is lower than the preset threshold value, the corresponding connection is marked as a to-be-removed state, and a preliminary list of removed connections is obtained.

[0022] Through the preliminary list of removed connections, a connection deletion operation is performed on the first network structure by using a matrix operation tool, and a temporary network structure after pruning is obtained.

[0023] For the temporary network structure, feature data of the vein texture is extracted through a texture analysis technique, the integrity of feature extraction is judged, if the integrity meets a preset condition, connection evaluation data is extracted from the temporary network structure.

[0024] According to the connection evaluation data, the effectiveness of the connection is classified by using a support vector machine, and an optimized connection distribution is obtained.

[0025] Through the optimized connection distribution, a final adjustment is performed on the temporary network structure, and a second network structure after pruning is obtained.

[0026] Preferably, the process of obtaining the third network structure with preliminary recognition capability comprises:

[0027] The feature representation of the palm vein image is extracted through a pre-established teacher network, and an initial feature set is obtained.

[0028] The knowledge distillation method is used to migrate features from the initial feature set to the second network structure, and a migration feature set is obtained.

[0029] The parameters of the second network structure are adjusted according to the migration feature set, and an optimized network parameter is obtained.

[0030] The second network structure is trained through the optimized network parameter, and a second network structure with preliminary recognition capability is obtained.

[0031] The output result of the second network structure is obtained, if the recognition capability is lower than a preset threshold value, the adjustment parameter is updated through a feedback mechanism, and an updated third network structure with preliminary recognition capability is obtained.

[0032] Preferably, the process of obtaining the quantized fourth network structure includes:

[0033] The sensitivity distribution is obtained by analyzing the response of each layer in the third network to low-contrast palm vein images through convolutional neural networks.

[0034] According to the sensitivity distribution, the sensitive layer and the non-sensitive layer are judged, and the layer allocation result is determined;

[0035] According to the layer allocation result, a mixed precision quantization strategy is adopted to allocate a 16-bit wide representation to the sensitive layer to obtain a first quantization parameter; and an 8-bit wide representation is allocated to the non-sensitive layer to obtain a second quantization parameter;

[0036] Fusion generates a fourth network structure based on the first quantization parameter and the second quantization parameter;

[0037] If there is precision loss in the fourth network structure, the quantization parameters are adjusted using a gradient descent algorithm to obtain a quantized fourth network structure.

[0038] Preferably, the process of obtaining the fifth network structure containing multi-scale information includes:

[0039] The palm vein image data is subjected to feature extraction through a multi-level module to obtain a preliminary feature set;

[0040] Using a feature pyramid network to process the preliminary feature set, capture features of different resolutions, and generate a multi-scale feature map;

[0041] Separating features of main blood vessels from the multi-scale feature map to obtain a first blood vessel feature subset;

[0042] Analyze the secondary branches according to the first vascular feature subset to obtain a second vascular feature subset;

[0043] detecting a capillary network using the second vascular feature subset to determine a third vascular feature subset;

[0044] Obtaining fusion information of the third blood vessel feature subset and the multi-scale feature map, and determining output features of the fifth network structure;

[0045] The network structure is optimized according to the output features of the fifth network structure to obtain the final multi-scale vascular network representation.

[0046] Preferably, the process of obtaining the optimized sixth network structure includes:

[0047] Obtain hierarchical feature data based on the venous network structure image and extract venous depth information;

[0048] Using a convolutional neural network to extract features from the venous network structure image to obtain a multi-layer feature map;

[0049] By calculating the correlation coefficient between each layer of feature map and vein depth information, a depth correlation matrix is ​​constructed;

[0050] If the correlation coefficient is greater than a preset threshold, the weight of the corresponding layer feature map is increased;

[0051] If the correlation coefficient is less than a preset threshold, the weight of the corresponding layer feature map is reduced;

[0052] According to the adjusted feature map weights, an attention mechanism module is designed to generate an attention map;

[0053] Performing weighted fusion of the attention map and the original feature map to obtain an enhanced feature representation;

[0054] Based on the enhanced feature representation, a fully connected layer is used to perform vein depth classification and output the optimized sixth network structure.

[0055] Preferably, the process of obtaining the seventh network structure for accurately extracting hierarchical features includes:

[0056] The output features are obtained through the sixth network, and the spatial distribution of the main blood vessels, secondary branches and capillaries is extracted using convolution operations to obtain the initial hierarchical representation;

[0057] For the initial hierarchical representation, a loss function is constructed to calculate the representation errors of the main blood vessels, secondary branches, and capillaries, and determine the error distribution;

[0058] If the error distribution exceeds a preset threshold, the network weights are adjusted through back propagation to obtain updated weight parameters;

[0059] The updated weight parameters are used to optimize the sixth network structure, generate the seventh network, and extract the accurate hierarchical representation;

[0060] According to the hierarchical representation of the seventh network, the feature consistency of the main blood vessels, secondary branches and capillaries is judged to obtain the feature extraction results;

[0061] Obtaining a hierarchical description of the vascular structure based on the feature extraction results, and determining a final seventh network output;

[0062] The optimized feature distribution is extracted from the final seventh network output to obtain an accurate hierarchical representation.

[0063] Preferably, the process of obtaining the final eighth network structure includes:

[0064] Obtaining the running data of the seventh network structure on the embedded device, including computing resource usage, processing speed and recognition accuracy;

[0065] Preprocessing the low-contrast palm vein image using an adaptive feature enhancement module to extract texture detail information from the image, and using a histogram equalization algorithm to enhance the contrast of the texture detail information to generate an enhanced palm vein feature image;

[0066] According to the preset threshold, it is determined whether the image features of the enhanced palm vein feature image meet the recognition accuracy requirements. If not, the adaptive feature enhancement module is returned to adjust the parameters;

[0067] For features that meet the accuracy requirements, a convolutional neural network is used to extract deep features, construct a palm vein feature vector, and a support vector machine classifier is used to classify and identify the feature vector to obtain the palm vein identity recognition result;

[0068] According to the recognition results and operation data, the network structure is optimized and adjusted to obtain the final eighth network structure.

[0069] The present invention also provides an embedded palm vein image recognition system based on the YOLO lightweight model, comprising:

[0070] The network simplification module is used to obtain the network structure of the original YOLO model. By analyzing the contribution of each convolutional layer to palm vein recognition, the number of redundant layers and filters is determined, and the standard convolution is replaced by depthwise separable convolution to obtain the simplified first network structure.

[0071] a pruning optimization module for calculating the importance of each connection to vein texture feature extraction based on the first network structure using a pruning technique, and removing the corresponding connection if the importance is lower than a preset threshold to obtain a pruned second network structure;

[0072] A knowledge distillation module is used to extract feature representations of palm vein images from a pre-established large-scale teacher network, transfer the features to the second network structure through a knowledge distillation method, and adjust the parameters of the student network to obtain a third network structure with preliminary recognition capabilities;

[0073] a hybrid quantization module, configured to analyze the sensitivity of each layer in the third network structure to low-contrast palm vein images, and adopt a mixed-precision quantization strategy to assign a 16-bit wide representation to the sensitive layer and an 8-bit wide representation to the non-sensitive layer, thereby obtaining a quantized fourth network structure;

[0074] a multi-scale feature module, configured to capture features of the main blood vessels, secondary branches, and capillary networks at different resolutions by designing a multi-level feature extraction module and adding a feature pyramid network to the fourth network structure, thereby obtaining a fifth network structure containing multi-scale information;

[0075] an attention enhancement module, configured to introduce an anatomically guided attention mechanism based on the vein layer features in the fifth network structure, calculate the correlation between the feature maps of each layer and the vein depth, obtain the model's attention to veins of different depths, and obtain an optimized sixth network structure;

[0076] a loss optimization module, configured to design a loss function based on the output features of the sixth network structure, adjust network weights by measuring the hierarchical representation errors of the main blood vessels, secondary branches, and capillary features, and obtain a seventh network structure that accurately extracts hierarchical features;

[0077] An adaptive enhancement module is configured to obtain operating data of the seventh network structure on an embedded device, amplify texture details of a low-contrast palm vein image using an adaptive feature enhancement module, determine whether the enhanced features meet recognition accuracy requirements, and obtain a final eighth network structure; and perform palm vein image recognition based on the eighth network structure to obtain an image recognition result.

[0078] Compared with the prior art, the present invention has the following advantages and technical effects:

[0079] This paper first analyzes the contribution of each layer of the YOLO model to palm vein recognition and streamlines the network structure using depthwise separable convolution and pruning techniques. It then uses knowledge distillation to transfer features from a large teacher network and introduces a mixed-precision quantization strategy to adapt to low-contrast images. Through a multi-level feature extraction module and an anatomically guided attention mechanism, the present invention enhances the recognition of veins at different depths. Finally, the hierarchical feature representation is optimized using a specific loss function, and its performance is verified on an embedded device. This method significantly improves the accuracy and efficiency of palm vein recognition while reducing computational resource requirements, providing an effective solution for biometric recognition on embedded devices.

[0080] The present invention realizes the lightweight, efficient and precise palm vein recognition model, significantly improves the recognition accuracy and operation efficiency, and is particularly suitable for resource-constrained embedded devices.

[0081] This method implements adaptive feature enhancement on embedded devices, effectively improving the recognition accuracy of low-contrast palm vein images. This method significantly enhances the performance and efficiency of palm vein recognition systems in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0083] Figure 1 Schematic diagram of a method flow in an embodiment of the present invention;

[0084] Figure 2 Schematic diagram of the system structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0085] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0086] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0087] Example 1

[0088] like Figure 1 As shown, this embodiment provides an embedded palm vein image recognition method based on the YOLO lightweight model, including:

[0089] The network structure of the original YOLO model was obtained. By analyzing the contribution of each convolutional layer to palm vein recognition, the number of redundant layers and filters was determined. Depthwise separable convolution was used to replace standard convolution to obtain the streamlined first network structure.

[0090] Based on the first network structure, the importance of each connection to the vein texture feature extraction is calculated using pruning technology. If the importance is lower than a preset threshold, the corresponding connection is removed to obtain a pruned second network structure;

[0091] Extract the feature representation of the palm vein image from the pre-established teacher network, transfer the features to the second network structure through the knowledge distillation method, adjust the parameters of the student network, and obtain the third network structure with preliminary recognition capabilities;

[0092] The sensitivity of each layer in the third network structure to low-contrast palm vein images was analyzed, and a mixed-precision quantization strategy was used to assign 16-bit width to the sensitive layers and 8-bit width to the non-sensitive layers, resulting in the quantized fourth network structure.

[0093] By designing a multi-level feature extraction module and adding a feature pyramid network to the fourth network structure, the features of the main blood vessels, secondary branches, and capillary networks at different resolutions are captured, resulting in a fifth network structure that contains multi-scale information.

[0094] Based on the vein layer features in the fifth network structure, an anatomically guided attention mechanism is introduced. By calculating the correlation between the feature maps of each layer and the vein depth, the model's attention to veins of different depths is obtained, resulting in the optimized sixth network structure.

[0095] A loss function is designed based on the output features of the sixth network structure. By measuring the hierarchical representation error of the main blood vessels, secondary branches, and capillary features, the network weights are adjusted to obtain the seventh network structure that accurately extracts hierarchical features.

[0096] The running data of the seventh network structure on the embedded device is obtained, and the texture details of the low-contrast palm vein image are amplified through the adaptive feature enhancement module. It is determined whether the enhanced features meet the recognition accuracy requirements to obtain the final eighth network structure; palm vein image recognition is performed based on the eighth network structure to obtain image recognition results.

[0097] Furthermore, the process of obtaining the simplified first network structure includes:

[0098] Obtain the network structure of the original YOLO model, decompose the convolutional layers to obtain the parameters and output features of each layer, and determine the preliminary analysis results;

[0099] By calculating the contribution of each convolutional layer output feature to palm vein recognition, redundant layers and unnecessary filters are determined to obtain a redundant judgment set;

[0100] Statistical methods are used to analyze the filter usage rate of the redundant decision set to determine the number of redundant layers and filters to be removed, and obtain a streamlined layer configuration;

[0101] Extract the retained convolutional layers from the streamlined layer configuration, and replace the standard convolution with depth-wise separable convolution to obtain the converted first network structure;

[0102] Run the palm vein dataset through the first network structure to obtain the recognition rate and computational complexity, determine whether it meets the preset threshold, and obtain the performance evaluation result;

[0103] If the performance evaluation result is lower than the preset threshold, the depthwise separable convolution parameters are adjusted to obtain the optimized network structure; otherwise, it is directly used as the first network structure.

[0104] Specifically, the original YOLO model's network structure typically consists of multiple convolutional layers for feature extraction and object detection. In the palm vein recognition scenario, the parameters and output features of each convolutional layer can be decomposed by analyzing its network structure.

[0105] For example, suppose the original YOLO model contains 20 convolutional layers, each with a different number of filters. For example, the first layer has 32 filters, and the output feature map is 416 × 416 × 32. By analyzing each layer, we can find that the output features of some layers contribute less to distinguishing palm vein texture. Therefore, the model structure can be simplified without affecting the recognition results.

[0106] Specifically, when calculating the contribution of the output features of each convolutional layer to palm vein recognition, this embodiment adopts a feature visualization method to observe whether the feature map of each layer effectively captures the texture details of the palm vein.

[0107] For example, if the feature map of layer 5 shows a large amount of low-frequency redundant information, while layer 10 can highlight the edges of the palm veins, then layer 5 may be determined to be a redundant layer.

[0108] Preferably, by counting the activation frequencies of the filters, assuming that only 10 of the 32 filters in the fifth layer are frequently activated, the remaining 22 filters can be regarded as non-essential filters to form a redundant decision set.

[0109] In one possible implementation, filter usage is analyzed for redundant decision sets. The average activation value of each layer's filters on the training set can be calculated. For example, if 40 of the 64 filters in layer 8 have activation values ​​below 0.1, these 40 filters are removed, leaving only 24, resulting in a streamlined layer configuration. This approach effectively reduces computational effort while preserving recognition capabilities.

[0110] It should be noted that the number of layers after streamlining may be reduced from 20 to 15, and the total number of filters is reduced from 512 to 300. After extracting the retained convolutional layers from the streamlined layer configuration, the standard convolution is replaced with depthwise separable convolution to form the first network structure. For example, the original 3×3 standard convolution requires 9 multiplication and addition operations, while the depthwise separable convolution is split into 3×3 depthwise convolution and 1×1 pointwise convolution, reducing the total number of multiplications and additions to approximately 3. This transformation significantly reduces computational complexity while maintaining feature extraction capabilities.

[0111] In one embodiment, the palm vein dataset is processed by the first network structure. Assuming that the recognition rate reaches 95%, the computational complexity is reduced from 100 GFLOPs to 30 GFLOPs. If the preset thresholds are 90% recognition rate and 50 GFLOPs, the performance meets the requirements.

[0112] It is understandable that if the recognition rate is only 85%, which is lower than the threshold, the depthwise separable convolution parameters are adjusted.

[0113] For example, increasing the number of channels in the depthwise convolution from 16 to 24 improved the recognition rate to 92%, resulting in an optimized network structure. This adjustment enhances the model's ability to perceive palm vein details. Ultimately, processing input data with the optimized network structure results in a streamlined palm vein recognition network. This network improves computational efficiency by approximately 70%, maintaining a high recognition rate and making it suitable for deployment in embedded devices.

[0114] For example, in practical applications, the streamlined network can achieve palm vein recognition at 10 frames per second on low-power devices, significantly improving efficiency compared to the original YOLO model's 2 frames per second. This approach not only reduces hardware requirements but also extends device battery life, demonstrating the practical value of technological optimization.

[0115] Furthermore, the process of obtaining the pruned second network structure includes:

[0116] Obtaining weight data of each connection in the first network structure through pruning technology, and calculating the contribution of each connection to the vein texture using a convolutional neural network to obtain an importance value;

[0117] Compare the importance value with a preset threshold. If the importance value is lower than the preset threshold, mark the corresponding connection as a state to be removed, and obtain a preliminary list of removed connections.

[0118] By removing the preliminary list of connections, a matrix operation tool is used to perform a connection deletion operation on the first network structure to obtain a temporary network structure after pruning;

[0119] For the temporary network structure, the feature data of the vein texture is extracted through texture analysis technology, and the integrity of the feature extraction is judged. If the integrity meets the preset conditions, the connection evaluation data is extracted from the temporary network structure;

[0120] Based on the connection evaluation data, support vector machine is used to classify the effectiveness of the connection and obtain the optimized connection distribution;

[0121] Through the optimized connection distribution, the temporary network structure is finally adjusted to obtain the pruned second network structure.

[0122] Specifically, based on the first network structure, weight data of each connection is obtained through pruning technology. It can be understood that the weight data reflects the degree of influence of each connection in the network on the result.

[0123] For example, in a palm vein recognition scenario, assume that the first network contains 1,000 connections. After analysis by the pruning tool, each connection is assigned a weight value ranging from 0 to 1. Connections with high weights may correspond to key features of texture edges, while connections with low weights may only process noise or redundant information.

[0124] In one possible implementation, a convolutional neural network (CNN) calculates each connection's contribution to vein texture. By combining forward and backpropagation, the sensitivity of each connection to the output feature can be extracted. For example, a connection with a weight of 0.8 might significantly influence vein boundary detection, while another connection with a weight of 0.2 might only have a slight effect on smooth areas. This importance value is generated and used for subsequent screening.

[0125] Specifically, when comparing the importance value to a preset threshold, this embodiment sets the threshold to 0.3. If a connection's importance value is 0.25, it is marked for removal. This approach ensures that connections that contribute substantially to texture recognition are retained. For example, a preliminary list might contain 300 low-importance connections, many of which are distributed in shallow convolution layers and may only capture low-level features.

[0126] It's important to note that when removing connections using matrix operation tools, the network can be treated as a sparse matrix. For example, if the original network had 5 million parameters, removing 300 connections would reduce the number to 4.8 million. This creates a temporary network structure, significantly reducing the computational burden.

[0127] In one embodiment, when determining feature extraction completeness through texture analysis, the main and branch features of the vein texture can be extracted. Assuming that completeness requires at least 95% retention of main features and at least 80% retention of branch features, if the temporary network meets these conditions, it is confirmed as the second network. This method ensures the preservation of core information.

[0128] For example, when extracting connection evaluation data from the second network, the density of connections in each layer can be calculated. For example, if a layer originally had 200 connections, and after pruning, only 150 remained, the support vector machine could classify them into "efficient" and "inefficient" categories, optimizing the connection distribution. Efficient connections might be concentrated in areas with texture detail, while inefficient connections are further eliminated.

[0129] In a possible implementation, when the second network is finally adjusted, the connection weights may be fine-tuned.

[0130] For example, the weight of a key connection in a certain layer was increased from 0.7 to 0.85 to enhance its response to texture features. This resulted in a stable network structure, improving both recognition efficiency and feature expression capabilities, contributing to the stable operation of subsequent palm vein recognition tasks.

[0131] Furthermore, the process of obtaining the third network structure with preliminary recognition capability includes:

[0132] The feature representation of the palm vein image is extracted through the pre-established teacher network to obtain the initial feature set;

[0133] Using the knowledge distillation method, features are transferred from the initial feature set to the second network structure to obtain a transferred feature set;

[0134] Adjust the parameters of the second network structure according to the migration feature set to obtain optimized network parameters;

[0135] The second network structure is trained by optimizing the network parameters to obtain a second network structure with preliminary recognition capability;

[0136] The output result of the second network structure is obtained. If the recognition ability is lower than a preset threshold, the adjustment parameters are updated through the feedback mechanism to obtain an updated third network structure with preliminary recognition ability.

[0137] Specifically, for the palm vein recognition task, the teacher network in this embodiment is a pre-trained, large-scale deep learning model. For example, this embodiment can use ResNet-152 as the teacher network. This network, pre-trained on the ImageNet dataset, possesses powerful feature extraction capabilities. By processing palm vein images with this teacher network, a rich feature representation can be obtained, including low-level texture features and high-level semantic features.

[0138] Knowledge distillation, a model compression technique, can transfer the knowledge of a teacher network to a smaller student network. Specifically, KL divergence can be used as the distillation loss function to guide the student network to learn the output distribution of the teacher network. In one possible implementation, the temperature parameter is set to 3 and the soft label weight is 0.7, which helps balance the contribution of distilled knowledge and true labels.

[0139] Network parameter optimization is a key step in improving model performance. For example, this embodiment can use the Adam optimizer, with the learning rate initially set to 0.001 and dynamically adjusted using a cosine annealing strategy. At the same time, L2 regularization is introduced with a regularization coefficient set to 0.0001 to help alleviate overfitting. During training, cross-validation can be used to evaluate the model's recognition capabilities.

[0140] It should be noted that if the recognition accuracy is lower than the preset threshold of 95%, the feedback mechanism will be triggered.

[0141] In one embodiment, a gradient clipping technique is used to limit the gradient value to the range of [-1, 1], which helps to stabilize the training process.

[0142] Preferably, after the model training is completed, the test set can be used to evaluate the final recognition performance. For example, the ROC curve and AUC indicators can be used to comprehensively evaluate the recognition ability of the model.

[0143] It is understandable that a high-quality recognition model can not only accurately identify known palm vein images, but also has good generalization capabilities for newly acquired image data, providing reliable technical support for practical applications.

[0144] Furthermore, the process of obtaining the quantized fourth network structure includes:

[0145] The sensitivity distribution is obtained by analyzing the response of each layer in the third network to low-contrast palm vein images through convolutional neural networks.

[0146] According to the sensitivity distribution, the sensitive layer and the non-sensitive layer are judged and the layer allocation result is determined;

[0147] According to the layer allocation result, a mixed precision quantization strategy is adopted to allocate 16-bit wide representation to the sensitive layer to obtain the first quantization parameter; and 8-bit wide representation is allocated to the non-sensitive layer to obtain the second quantization parameter;

[0148] Fusion generates a fourth network structure based on the first quantization parameter and the second quantization parameter;

[0149] If there is precision loss in the fourth network structure, the quantization parameters are adjusted using a gradient descent algorithm to obtain a quantized fourth network structure.

[0150] Specifically, the response of each layer in the third network to low-contrast palm vein images is analyzed through convolutional neural networks, and obtaining the sensitivity distribution is a key link.

[0151] For example, by inputting a set of low-contrast palm vein images, we can observe the activation of each convolution kernel layer, noting which layers are more responsive to detailed textures and which layers have smoother outputs. Assuming the third network has 10 convolutional layers, analysis reveals that the first three layers are likely more sensitive to image edges and subtle textures, while the later layers tend to extract global features. This distribution reflects the network's processing characteristics for low-contrast data and can aid subsequent optimization.

[0152] In one possible implementation, a threshold can be set when determining sensitive layers and non-sensitive layers. For example, layers with an activation value mean exceeding 0.7 are classified as sensitive layers, and the rest are non-sensitive layers.

[0153] For example, assuming the mean activation values ​​of layers 1 and 2 are 0.8 and 0.75, respectively, while that of layer 5 is only 0.3, the first two layers are classified as sensitive layers. This logical division provides a basis for subsequent quantization. Using a mixed-precision quantization strategy for layer allocation results is a highly efficient resource allocation method.

[0154] Specifically, assigning 16-bit wide representation to the sensitive layer can retain more detailed features.

[0155] For example, when processing low-contrast palm vein images, the first layer may need to capture subtle texture changes, and 16-bit precision ensures that this information is not lost. Using 8-bit width for the insensitive layer reduces computational effort. Assuming the fifth layer primarily extracts large-scale contours, 8 bits are sufficient. This differentiated allocation ensures both accuracy and efficiency. After obtaining the first and second quantization parameters, fusing them to generate the fourth network structure is a crucial step.

[0156] In an embodiment, the 16-bit and 8-bit feature maps can be integrated by parameter splicing.

[0157] For example, the output of the sensitive layer is adjusted to a uniform scale, and then spliced with the output of the non-sensitive layer to generate a new feature set. This method can maintain the integrity of the information.

[0158] It should be noted that if the accuracy decreases after fusion, such as the recognition rate decreases from 90% to 85%, further adjustment is required. If the fourth network structure has precision loss, adjusting the quantization parameters through gradient descent algorithm is a feasible solution. For example, the quantization step can be gradually reduced from 0.01 to 0.005 for the quantization parameters of the sensitive layer, and whether the accuracy increases is observed.

[0159] In an embodiment, it is found that the sensitive layer details are more complete after adjustment, and the recognition rate is increased to 88%. This shows that fine-tuning the quantization parameters can effectively compensate for the loss. The final goal is to verify the feature extraction capability of the low-contrast palm vein image through the optimized network structure.

[0160] It can be understood that a set of test images, such as 50 low-contrast samples, are input to observe whether the network can correctly extract key texture features.

[0161] For example, the palm vein pattern in an image is blurred, and the optimized network can still identify the main branches, while the unoptimized network may only output noise. This improvement in capability is of great significance to practical applications.

[0162] Specifically, the implementation of the final quantized fourth network structure can be regarded as a balance between resources and accuracy. For example, when deployed on an embedded device, the 8-bit non-sensitive layer reduces the memory usage by about 30%, while the 16-bit sensitive layer ensures that key features are not lost. This design not only reduces power consumption, but also maintains recognition effectiveness, and is particularly suitable for palm vein recognition scenarios that are sensitive to details.

[0163] Further, the process of obtaining the fifth network structure containing multi-scale information includes:

[0164] The palm vein image data is processed by a multi-level module to obtain a preliminary feature set;

[0165] The preliminary feature set is processed by a feature pyramid network to capture different resolution features and generate a multi-scale feature map;

[0166] The features of the main blood vessels are separated from the multi-scale feature map to obtain a first blood vessel feature subset;

[0167] The secondary branches are analyzed based on the first blood vessel feature subset to obtain a second blood vessel feature subset;

[0168] detecting a capillary network by using the second vascular feature subset to determine a third vascular feature subset;

[0169] Obtaining fusion information of the third vascular feature subset and the multi-scale feature map to determine the output features of the fifth network structure;

[0170] The network structure is optimized according to the output features of the fifth network structure to obtain the final multi-scale vascular network representation.

[0171] Specifically, the input data is subjected to feature extraction through a multi-level module to obtain a preliminary feature set. This process can be understood as decomposing the original low-contrast palm vein image into multiple basic features.

[0172] For example, in one possible implementation, edge information and texture details can be extracted through convolution operations. Assuming the input image resolution is 512×512 and the pixel value range is between 0-255, the multi-level module may contain three layers of convolution, each layer using convolution kernels of different sizes, such as 3×3 and 5×5, to capture local features of different ranges. This hierarchical extraction helps preserve key structural information in the image.

[0173] The feature pyramid network is used to process the initial feature set, capture features at different resolutions, and generate multi-scale feature maps. The core of this design is to address the problem of insufficient representation of single-resolution features. Specifically, the feature pyramid network can combine high-level semantic information with low-level detailed information.

[0174] For example, a low-level feature map might maintain a 256×256 resolution to capture subtle textures, while a high-level feature map might be scaled down to 64×64 to extract more abstract structural features. By fusing upsampling and downsampling, a multi-scale feature map can simultaneously capture the details of both the trunk and branches of the palmar vein. Separating the features of the primary vessels from the multi-scale feature map to obtain the first subset of vessel features requires targeted analysis of the feature maps.

[0175] In one embodiment, threshold segmentation techniques can be used to distinguish major vessel regions. Assuming that the pixel intensity of major vessels is 1.5 times greater than the average, morphological operations are combined to remove noise points, thereby extracting a clear outline of the main vessels. This method has the advantage of quickly locating core structures in the image. Secondary branches are analyzed within the first vessel feature subset to obtain a second vessel feature subset, further refining the main vessels.

[0176] Preferably, a directional filter can be introduced to analyze the direction of vascular extension, for example, using a set of predefined angle templates to match the direction of secondary branches. By comparing the response strength in different directions, secondary branch features can be screened. This refinement helps improve the description of complex vascular networks. The capillary network is detected using the second vascular feature subset, and the third vascular feature subset is determined. This stage focuses on the identification of microstructures.

[0177] It is understandable that capillaries are difficult to extract directly due to their low contrast, so we can combine high-resolution feature maps and use local enhancement technology to amplify weak signals.

[0178] For example, by amplifying areas below a certain threshold in a feature map and then using cluster analysis to determine the distribution of capillaries, this approach can effectively supplement the network's perception of subtle features.

[0179] The fusion information of the third vascular feature subset and the multi-scale feature map is obtained to determine the output features of the fifth network structure. This process aims to integrate multi-level information. In one possible implementation, the third vascular feature subset and the multi-scale feature map can be added element-by-element through a weighted fusion method. The weights are dynamically adjusted based on the feature importance, for example, the weight of the main vessel is set to 0.6 and the weight of the capillaries is set to 0.3. This fusion can generate a more comprehensive feature representation. The network structure is optimized based on the output features of the fifth network structure to obtain the final multi-scale vascular network representation. This optimization process reflects the network's adaptability. For example, the fusion weights can be adjusted through backpropagation to make the output feature map more closely resemble the actual vascular distribution.

[0180] For example, if the initial feature map is broken at the branch connection, the network's learning of continuity can be strengthened by increasing the number of training iterations or adjusting the loss function. The benefit of this optimization is to improve the robustness and completeness of feature extraction, providing reliable support for subsequent applications.

[0181] Furthermore, the process of obtaining the optimized sixth network structure includes:

[0182] Obtain hierarchical feature data based on the venous network structure image and extract venous depth information;

[0183] A convolutional neural network is used to extract features from the venous network structure image to obtain a multi-layer feature map;

[0184] By calculating the correlation coefficient between each layer of feature map and vein depth information, a depth correlation matrix is ​​constructed;

[0185] If the correlation coefficient is greater than the preset threshold, the weight of the corresponding layer feature map is increased;

[0186] If the correlation coefficient is less than the preset threshold, the weight of the corresponding layer feature map is reduced;

[0187] According to the adjusted feature map weights, an attention mechanism module is designed to generate an attention map;

[0188] Perform weighted fusion of the attention map and the original feature map to obtain enhanced feature representation;

[0189] Based on the enhanced feature representation, a fully connected layer is used to perform vein depth classification and output the optimized sixth network structure.

[0190] Specifically, convolutional neural networks are a core tool for obtaining hierarchical feature data from venous network structure images and extracting venous depth information. In one possible implementation, multi-layer convolution operations can be used to extract basic features such as edges and textures from the input venous image, gradually transitioning to more abstract deep features. Assuming that the input image is a grayscale image with a pixel value range of 0-255, shallow convolution may focus on the grayscale changes of the vein edges, while deep convolution captures the overall pattern of vein distribution. This hierarchical extraction method can gradually build feature representations from local to global, which is helpful for subsequent analysis of depth information.

[0191] Specifically, the correlation coefficient reflects the degree of match between the feature map and the depth information. For example, assuming there are three layers of feature maps, the first layer captures vein edges with a correlation coefficient of 0.8; the second layer captures vein branches with a coefficient of 0.6; and the third layer captures the overall structure with a coefficient of only 0.3. If the preset threshold is 0.5, the weights of the first and second layers will increase, while the weight of the third layer will decrease. This method uses data-driven adjustment of feature importance to ensure that depth information relies more on strongly correlated features.

[0192] In one embodiment, spatial attention can be incorporated into the design of the attention mechanism module based on the adjusted feature map weights. The attention map can highlight key areas of vein depth. For example, for a deeper vein, the attention map might generate a higher weight value, such as 0.9, in the corresponding area, while a lower weight value, such as 0.2, is generated for shallower areas. This generated attention map can effectively guide the network to focus on depth-related areas, thereby improving feature quality.

[0193] It's important to note that the weighted fusion of the attention map and the original feature map enhances the feature representation. For example, during fusion, the weights of the attention map can be element-wise multiplied with the feature map, amplifying the feature values ​​of deep vein regions while suppressing irrelevant areas. This fusion approach allows the enhanced feature representation to focus more on depth information, providing a more reliable foundation for subsequent classification. Based on this enhanced feature representation, vein depth classification using a fully connected layer is the final step in output optimization.

[0194] In one possible implementation, the fully connected layer can be designed to have three types of outputs: shallow, middle and deep veins. For example, assuming that an image containing multiple veins is input, the network can output that the proportion of shallow veins is 40%, the proportion of middle veins is 30%, and the proportion of deep veins is 30%. This classification result not only clearly reflects the distribution of vein depth, but also provides intuitive data support for subsequent applications.

[0195] Specifically, the implementation of the attention mechanism can be further expanded from multiple aspects.

[0196] For example, the channel attention can be combined to adjust the weights of different feature channels respectively, highlighting the feature dimensions most relevant to the depth. Preferably, if a channel is strongly related to deep veins, its weight can be increased to 0.85, while the weight of a weakly related channel can be reduced to 0.1. This multi-dimensional attention design can optimize the feature representation from both spatial and channel aspects.

[0197] It can be understood that each step of the above method is closely related to the extraction of vein depth information.

[0198] For example, the calculation of the correlation coefficient ensures the scientificity of feature selection, the introduction of the attention mechanism improves the pertinence of the features, and the weighted fusion and fully connected classification ensure the accuracy of the output. These links support the complete process from image to depth classification through the logical relationship of layer-by-layer progression.

[0199] In one embodiment, if the input image resolution is 512x512, the network can effectively distinguish vein regions of different depths after processing, providing high-quality basic data for subsequent analysis.

[0200] Further, the process of obtaining the seventh network structure for accurately extracting hierarchical features includes:

[0201] The output features are obtained by the sixth network, and the convolution operation is used to extract the spatial distribution of the main blood vessels, secondary branches and capillaries to obtain the initial hierarchical representation;

[0202] For the initial hierarchical representation, a loss function is constructed to calculate the representation error of the main blood vessels, secondary branches and capillaries, and determine the error distribution;

[0203] If the error distribution exceeds a preset threshold, the network weights are adjusted through back propagation to obtain updated weight parameters;

[0204] The updated weight parameters are used to optimize the sixth network structure to generate the seventh network, which extracts accurate hierarchical representation;

[0205] According to the hierarchical representation of the seventh network, the consistency of the features of the main blood vessels, secondary branches and capillaries is judged to obtain a feature extraction result;

[0206] According to the feature extraction result, a hierarchical description of the blood vessel structure is obtained, and a final output of the seventh network is determined;

[0207] From the final output of the seventh network, an optimized feature distribution is extracted to obtain an accurate hierarchical representation.

[0208] Specifically, for the output features of the sixth network, the spatial distribution of the main blood vessels, secondary branches and capillaries is extracted by convolution operation. This process can be understood as sliding window processing of the input image by the convolution kernel to capture blood vessel features of different scales.

[0209] For example, when processing a venous network image, the convolution kernel size can be set to 3x3 or 5x5 to extract the outline of thick blood vessels and the details of small branches, respectively.

[0210] For example, if the input image resolution is 256x256 and the pixel value range is 0-255, a multi-channel feature map can be generated after convolution operation, for example, with 64 channels, each channel highlighting the spatial characteristics of a certain type of blood vessel. This way can effectively preserve the spatial position information of the blood vessels and provide a basis for subsequent hierarchical representation. After constructing the initial hierarchical representation, a loss function is designed for the main blood vessels, secondary branches and capillaries to calculate the representation error. The core of this step is to quantify the accuracy of feature extraction.

[0211] Specifically, the difference between the predicted distribution and the true distribution can be measured by mean square error or cross-entropy loss.

[0212] For example, if the true distribution of the main blood vessels is 40%, the secondary branches are 35%, and the capillaries are 25%, and the model prediction values are 38%, 36% and 26% respectively, the error distribution can be obtained by comparing each item. This error calculation helps to identify the bias of the model for different types of blood vessels and provides direction for subsequent optimization.

[0213] It should be noted that if the error distribution exceeds a predetermined threshold, for example, 0.05, the network weights are adjusted by backpropagation. In one possible implementation, backpropagation can be based on gradient descent method, and the learning rate is set to 0.001 to update the convolution layer weights step by step.

[0214] For example, when the main blood vessel feature error is large, the weight of the related convolution kernel will be adjusted more significantly to enhance the attention to this type of feature. This dynamic adjustment can make the model more suitable for the distribution characteristics of different types of blood vessels.

[0215] Preferably, the process of generating the seventh network structure using the updated weight parameters optimizes the sixth network structure, similar to a structure iteration. For example, the original sixth network may have 10 convolutional layers, and the optimized seventh network may add a feature fusion layer, increasing the number of channels from 64 to 128, further integrating multi-scale features. This structure upgrade can improve the model's ability to represent complex vascular networks.

[0216] In one embodiment, when judging the consistency of the features of the main blood vessels, secondary branches and capillaries, the cosine similarity can be introduced as a measure. For example, after extracting the feature vector of the main blood vessels, compare it with the standard template vector. If the similarity is higher than 0.9, it is considered to be in good consistency. This method quantifies the relationship between features to ensure the reliability of hierarchical representation.

[0217] Specifically, by obtaining the hierarchical description of the vascular structure through the feature extraction result, when determining the final output, the feature map can be divided into three types of regions.

[0218] For example, in a 512x512 feature map, the pixel values of the main blood vessel region are concentrated in 200-255, the secondary branch is 100-199, and the capillary is 0-99. This hierarchical description intuitively reflects the spatial hierarchical distribution of blood vessels.

[0219] For example, when extracting the optimized feature distribution from the final output, the feature map can be observed through a visualization tool.

[0220] For example, in a certain experiment, the optimized feature map shows that the main blood vessel edge is sharper and the capillary details are richer. This accurate hierarchical representation can provide a higher quality data basis for subsequent analysis, which helps to improve the depth and breadth of vascular network research.

[0221] It can be understood that the whole process from feature extraction to optimized output forms a closed-loop iterative mechanism. The adjustment of each link is around improving the accuracy of vascular hierarchical representation, and the final generated feature distribution has significant advantages in spatial resolution and class distinction, laying a solid foundation for further research on venous network.

[0222] Further, the process of obtaining the final eighth network structure includes:

[0223] Obtain the running data of the seventh network structure on the embedded device, including the occupation of computing resources, processing speed and recognition accuracy;

[0224] The low-contrast palm vein image is preprocessed by the adaptive feature enhancement module, the texture detail information in the image is extracted, the histogram equalization algorithm is used to enhance the contrast of the texture detail information, and the enhanced palm vein feature image is generated;

[0225] According to the preset threshold, it is determined whether the image features of the enhanced palm vein feature image meet the recognition accuracy requirements. If not, the adaptive feature enhancement module is returned to adjust the parameters;

[0226] For features that meet the accuracy requirements, a convolutional neural network is used to extract deep features and construct a palm vein feature vector. The support vector machine classifier is used to classify and identify the feature vector to obtain the palm vein identity recognition result.

[0227] According to the recognition results and operation data, the network structure is optimized and adjusted to obtain the final eighth network structure.

[0228] Specifically, the adaptive feature enhancement module of this implementation analyzes the grayscale distribution of the input image and dynamically adjusts the enhancement parameters to effectively enhance the details of low-contrast areas.

[0229] For example, for a palm vein image with low contrast, the module may first apply an adaptive histogram equalization algorithm to redistribute the image grayscale value to the full range of 0-255 to highlight the vein texture.

[0230] It should be noted that histogram equalization may introduce noise. Therefore, in one possible implementation, the system will combine noise reduction methods such as Gaussian filtering to enhance contrast while maintaining image smoothness.

[0231] Specifically, a 3x3 or 5x5 Gaussian kernel can be used for convolution, and the σ value can be dynamically adjusted according to the degree of image noise, usually between 0.5 and 2. In the feature extraction stage, the design of the convolutional neural network is crucial.

[0232] In one embodiment, the network adopts a structure similar to VGG, including multiple convolutional layers, pooling layers and fully connected layers.

[0233] Preferably, the first convolutional layer uses a larger convolution kernel (such as 7x7) to capture large-scale features, and subsequent layers gradually reduce the convolution kernel size to 3x3 to extract finer texture information.

[0234] It can be understood that this progressive feature extraction helps to construct multi-scale palm vein representation.

[0235] For example, for a 300x300 pixel input image, the network might consist of five convolutional blocks, each containing two to three convolutional layers and a max pooling layer. Finally, a feature vector of fixed length (e.g., 512 dimensions) is obtained through global average pooling. This design ensures feature discriminability while controlling computational complexity, making it suitable for deployment on embedded devices. During the classification phase, the choice of kernel function for the support vector machine (SVM) directly impacts recognition performance.

[0236] In one embodiment, a radial basis function (RBF) kernel may be used, with its parameter γ determined by cross-validation and typically in the range of 0.001-0.1. This nonlinear kernel function can effectively handle complex decision boundaries in high-dimensional feature spaces and improve recognition accuracy.

[0237] It should be further explained that the hardware platform construction of this embodiment includes:

[0238] The STM32F407VET6 is used as the control platform, responsible for the system's control logic and data transmission. This platform has high performance and low power consumption, making it suitable for controlling embedded devices.

[0239] Main control chip: Jetson Xavier NX is selected as the main control chip to run the palm vein image recognition algorithm. Jetson Xavier NX has powerful computing power and can efficiently run the lightweight YOLO model.

[0240] Camera module: Connect the camera module to capture palm vein images.

[0241] Specifically, the optimized YOLO lightweight model is deployed on the Jetson Xavier NX main control chip.

[0242] The model is run on the embedded device to perform real-time recognition on the collected palm vein images.

[0243] The recognition results are output to external devices such as display screens or communication modules through the STM32F407VET6 control platform.

[0244] This embodiment significantly reduces the model's computational complexity and storage requirements through a lightweight design, enabling it to run efficiently on resource-constrained embedded devices. By employing techniques such as multi-level feature extraction, an attention mechanism, and a specific loss function, the recognition accuracy of palm vein images is effectively improved, enabling accurate distinction of vein features at different depths. The optimized model can quickly complete palm vein image recognition on embedded devices, meeting real-time requirements and making it suitable for practical application scenarios. Techniques such as data augmentation and adaptive feature enhancement improve the model's adaptability to varying lighting conditions, palm postures, and low-contrast images, thereby enhancing the robustness of the system.

[0245] Example 2

[0246] like Figure 2 As shown, based on the same inventive concept, this embodiment also provides an embedded palm vein image recognition system based on a YOLO lightweight model, including:

[0247] The network simplification module is used to obtain the network structure of the original YOLO model. By analyzing the contribution of each convolutional layer to palm vein recognition, the number of redundant layers and filters is determined, and the standard convolution is replaced by depthwise separable convolution to obtain the simplified first network structure.

[0248] A pruning optimization module is used to calculate the importance of each connection to the vein texture feature extraction based on the first network structure using pruning technology. If the importance is lower than a preset threshold, the corresponding connection is removed to obtain a pruned second network structure;

[0249] The knowledge distillation module is used to extract the feature representation of palm vein images from a pre-established large-scale teacher network, transfer the features to the second network structure through the knowledge distillation method, and adjust the parameters of the student network to obtain a third network structure with preliminary recognition capabilities;

[0250] A hybrid quantization module is used to analyze the sensitivity of each layer in the third network structure to low-contrast palm vein images. A mixed-precision quantization strategy is used to assign 16-bit width representation to sensitive layers and 8-bit width representation to non-sensitive layers, thereby obtaining a quantized fourth network structure.

[0251] A multi-scale feature module is used to capture the features of the main blood vessels, secondary branches, and capillary networks at different resolutions by designing a multi-level feature extraction module and adding a feature pyramid network to the fourth network structure, thereby obtaining a fifth network structure containing multi-scale information;

[0252] The attention enhancement module is used to introduce an anatomically guided attention mechanism based on the vein layer features in the fifth network structure. By calculating the correlation between the feature maps of each layer and the vein depth, the model's attention to veins of different depths is obtained, resulting in an optimized sixth network structure.

[0253] a loss optimization module for designing a loss function based on the output features of the sixth network structure, adjusting the network weights by measuring the hierarchical representation errors of the main blood vessels, secondary branches, and capillary features, and obtaining a seventh network structure that accurately extracts hierarchical features;

[0254] The adaptive enhancement module is used to obtain the operating data of the seventh network structure on the embedded device, amplify the texture details of the low-contrast palm vein image through the adaptive feature enhancement module, determine whether the enhanced features meet the recognition accuracy requirements, and obtain the final eighth network structure; perform palm vein image recognition based on the eighth network structure to obtain image recognition results.

[0255] The embedded palm vein image recognition system based on the YOLO lightweight model provided in this embodiment has all the advantages of the embedded palm vein image recognition method based on the YOLO lightweight model provided in Example 1.

[0256] Example 3

[0257] This embodiment further discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in the first embodiment.

[0258] Example 4

[0259] This embodiment further discloses a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0260] Example 5

[0261] This embodiment further discloses a computer program product, including a computer program, which implements the steps of the method described in the first embodiment when executed by a processor.

[0262] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An embedded palm vein image recognition method based on the YOLO lightweight model, characterized in that: include: The network structure of the original YOLO model was obtained. By analyzing the contribution of each convolutional layer to palm vein recognition, the number of redundant layers and filters was determined. Depthwise separable convolution was used to replace standard convolution to obtain the streamlined first network structure. Based on the first network structure, the importance of each connection to the vein texture feature extraction is calculated using pruning technology. If the importance is lower than a preset threshold, the corresponding connection is removed to obtain a pruned second network structure; Extracting feature representations of palm vein images from a pre-established teacher network, migrating these features to the second network structure through knowledge distillation, and adjusting student network parameters to obtain a third network structure with preliminary recognition capabilities. Analyze the sensitivity of each layer in the third network structure to low-contrast palm vein images, adopt a mixed precision quantization strategy, assign a 16-bit wide representation to the sensitive layer and an 8-bit wide representation to the non-sensitive layer, and obtain a quantized fourth network structure; By designing a multi-level feature extraction module and adding a feature pyramid network to the fourth network structure, the features of the main blood vessels, secondary branches and capillary networks at different resolutions are captured, thereby obtaining a fifth network structure containing multi-scale information; Based on the vein layer features in the fifth network structure, an anatomically guided attention mechanism is introduced. By calculating the correlation between the feature maps of each layer and the vein depth, the model's attention to veins of different depths is obtained, resulting in an optimized sixth network structure. A loss function is designed based on the output features of the sixth network structure, and the network weights are adjusted by measuring the hierarchical representation errors of the main blood vessels, secondary branches, and capillary features to obtain a seventh network structure that accurately extracts hierarchical features; Obtaining running data of the seventh network structure on the embedded device, amplifying texture details of the low-contrast palm vein image using an adaptive feature enhancement module, determining whether the enhanced features meet recognition accuracy requirements, and obtaining a final eighth network structure; Performing palm vein image recognition based on the eighth network structure to obtain an image recognition result; The process of obtaining the quantized fourth network structure includes: The sensitivity distribution is obtained by analyzing the response of each layer in the third network to low-contrast palm vein images through convolutional neural networks. According to the sensitivity distribution, the sensitive layer and the non-sensitive layer are judged, and the layer allocation result is determined; According to the layer allocation result, a mixed precision quantization strategy is adopted to allocate a 16-bit wide representation to the sensitive layer to obtain a first quantization parameter; and an 8-bit wide representation is allocated to the non-sensitive layer to obtain a second quantization parameter; Fusion generates a fourth network structure based on the first quantization parameter and the second quantization parameter; If there is precision loss in the fourth network structure, adjusting the quantization parameters by a gradient descent algorithm to obtain a quantized fourth network structure; The process of obtaining the fifth network structure containing multi-scale information includes: The palm vein image data is subjected to feature extraction through a multi-level module to obtain a preliminary feature set; Using a feature pyramid network to process the preliminary feature set, capture features of different resolutions, and generate a multi-scale feature map; Separating features of main blood vessels from the multi-scale feature map to obtain a first blood vessel feature subset; Analyze the secondary branches according to the first vascular feature subset to obtain a second vascular feature subset; detecting a capillary network using the second vascular feature subset to determine a third vascular feature subset; Obtaining fusion information of the third blood vessel feature subset and the multi-scale feature map, and determining output features of the fifth network structure; Optimize the network structure according to the output characteristics of the fifth network structure to obtain the final multi-scale vascular network representation; The process of obtaining the optimized sixth network structure includes: Obtain hierarchical feature data based on the venous network structure image and extract venous depth information; Using a convolutional neural network to extract features from the venous network structure image to obtain a multi-layer feature map; By calculating the correlation coefficient between each layer of feature map and vein depth information, a depth correlation matrix is ​​constructed; If the correlation coefficient is greater than a preset threshold, the weight of the corresponding layer feature map is increased; If the correlation coefficient is less than a preset threshold, the weight of the corresponding layer feature map is reduced; According to the adjusted feature map weights, an attention mechanism module is designed to generate an attention map; Performing weighted fusion of the attention map and the original feature map to obtain an enhanced feature representation; Based on the enhanced feature representation, a fully connected layer is used to perform vein depth classification and output the optimized sixth network structure; The process of obtaining the seventh network structure for accurately extracting hierarchical features includes: The output features are obtained through the sixth network, and the spatial distribution of the main blood vessels, secondary branches and capillaries is extracted using convolution operations to obtain the initial hierarchical representation; For the initial hierarchical representation, a loss function is constructed to calculate the representation errors of the main blood vessels, secondary branches, and capillaries, and determine the error distribution; If the error distribution exceeds a preset threshold, the network weights are adjusted through back propagation to obtain updated weight parameters; The updated weight parameters are used to optimize the sixth network structure, generate the seventh network, and extract the accurate hierarchical representation; According to the hierarchical representation of the seventh network, the feature consistency of the main blood vessels, secondary branches and capillaries is judged to obtain the feature extraction results; Obtaining a hierarchical description of the vascular structure based on the feature extraction results, and determining a final seventh network output; Extract the optimized feature distribution from the final seventh network output to obtain an accurate hierarchical representation; The process of obtaining the final eighth network structure includes: Obtaining the running data of the seventh network structure on the embedded device, including computing resource usage, processing speed and recognition accuracy; Preprocessing the low-contrast palm vein image using an adaptive feature enhancement module to extract texture detail information from the image, and using a histogram equalization algorithm to enhance the contrast of the texture detail information to generate an enhanced palm vein feature image; According to the preset threshold, it is determined whether the image features of the enhanced palm vein feature image meet the recognition accuracy requirements. If not, the adaptive feature enhancement module is returned to adjust the parameters; For features that meet the accuracy requirements, a convolutional neural network is used to extract deep features, construct a palm vein feature vector, and a support vector machine classifier is used to classify and identify the feature vector to obtain the palm vein identity recognition result; According to the recognition results and operation data, the network structure is optimized and adjusted to obtain the final eighth network structure.

2. The method according to claim 1, characterized in that The process of obtaining the simplified first network structure includes: Obtain the network structure of the original YOLO model, decompose the convolutional layers to obtain the parameters and output features of each layer, and determine the preliminary analysis results; By calculating the contribution of each convolutional layer output feature to palm vein recognition, redundant layers and unnecessary filters are determined to obtain a redundant judgment set; Analyzing the filter usage rate of the redundant decision set using a statistical method, determining the number of redundant layers and filters to be removed, and obtaining a streamlined layer configuration; Extracting the retained convolutional layers from the streamlined layer configuration, replacing the standard convolution with depthwise separable convolution, to obtain a converted first network structure; Running the palm vein dataset through the first network structure, obtaining the recognition rate and computational complexity, determining whether the preset threshold is met, and obtaining a performance evaluation result; If the performance evaluation result is lower than the preset threshold, the depthwise separable convolution parameters are adjusted to obtain the optimized network structure; otherwise, it is directly used as the first network structure.

3. The method according to claim 1, characterized in that The process of obtaining the pruned second network structure includes: Obtaining weight data of each connection in the first network structure through pruning technology, and calculating the contribution of each connection to the vein texture using a convolutional neural network to obtain an importance value; Comparing the importance value with a preset threshold, if the importance value is lower than the preset threshold, marking the corresponding connection as a state to be removed, and obtaining a preliminary list of removed connections; Using the preliminary list of removed connections, a matrix operation tool is used to perform a connection deletion operation on the first network structure to obtain a temporary network structure after pruning; Extracting vein texture feature data from the temporary network structure using texture analysis technology, determining the integrity of the feature extraction, and extracting connection evaluation data from the temporary network structure if the integrity meets a preset condition; Based on the connection evaluation data, a support vector machine is used to classify the effectiveness of the connection to obtain an optimized connection distribution; The temporary network structure is finally adjusted by optimizing the connection distribution to obtain a pruned second network structure.

4. The method according to claim 1, wherein The process of obtaining a third network structure with preliminary recognition capabilities includes: The feature representation of the palm vein image is extracted through the pre-established teacher network to obtain the initial feature set; Using a knowledge distillation method, features are transferred from the initial feature set to a second network structure to obtain a transferred feature set; Adjusting the parameters of the second network structure according to the migration feature set to obtain optimized network parameters; The second network structure is trained by optimizing the network parameters to obtain a second network structure with preliminary recognition capability; The output result of the second network structure is obtained. If the recognition ability is lower than a preset threshold, the adjustment parameters are updated through the feedback mechanism to obtain an updated third network structure with preliminary recognition ability.

5. An embedded palm vein image recognition system based on a YOLO lightweight model, used to implement the method according to any one of claims 1 to 4, characterized in that: include: The network simplification module is used to obtain the network structure of the original YOLO model. By analyzing the contribution of each convolutional layer to palm vein recognition, the number of redundant layers and filters is determined, and the standard convolution is replaced by depthwise separable convolution to obtain the simplified first network structure. a pruning optimization module for calculating the importance of each connection to vein texture feature extraction based on the first network structure using a pruning technique, and removing the corresponding connection if the importance is lower than a preset threshold to obtain a pruned second network structure; A knowledge distillation module is used to extract feature representations of palm vein images from a pre-established large-scale teacher network, transfer the features to the second network structure through a knowledge distillation method, and adjust the parameters of the student network to obtain a third network structure with preliminary recognition capabilities; a hybrid quantization module, configured to analyze the sensitivity of each layer in the third network structure to low-contrast palm vein images, and adopt a mixed-precision quantization strategy to assign a 16-bit wide representation to the sensitive layer and an 8-bit wide representation to the non-sensitive layer, thereby obtaining a quantized fourth network structure; a multi-scale feature module, configured to capture features of the main blood vessels, secondary branches, and capillary networks at different resolutions by designing a multi-level feature extraction module and adding a feature pyramid network to the fourth network structure, thereby obtaining a fifth network structure containing multi-scale information; an attention enhancement module, configured to introduce an anatomically guided attention mechanism based on the vein layer features in the fifth network structure, calculate the correlation between the feature maps of each layer and the vein depth, obtain the model's attention to veins of different depths, and obtain an optimized sixth network structure; a loss optimization module, configured to design a loss function based on the output features of the sixth network structure, adjust network weights by measuring the hierarchical representation errors of the main blood vessels, secondary branches, and capillary features, and obtain a seventh network structure that accurately extracts hierarchical features; an adaptive enhancement module, configured to obtain operating data of the seventh network structure on an embedded device, amplify texture details of the low-contrast palm vein image using an adaptive feature enhancement module, determine whether the enhanced features meet recognition accuracy requirements, and obtain a final eighth network structure; Palm vein image recognition is performed based on the eighth network structure to obtain an image recognition result.

Images