Application method of SMA-ConvNeXt model in palmprint information personnel identity recognition

Abstract

The invention discloses an SMA-ConvNeXt model-based application method in palmprint information personnel identity recognition, which comprises the following steps of: making an SMPDRED infrared integrated palmprint information recognition benchmark test set, preprocessing a data set, and dividing the data set into a training set and a verification set; building a ConvNeXt basic model, initializing model parameters, and training the basic model by adopting a training set; the ConvNeXt basic model is improved, and an improved SMA-ConvNeXt model is obtained; the improved SMA-ConvNeXt model is verified, and model indexes of the improved SMA-ConvNeXt model and a ConvNeXt basic model are compared; and according to the improved SMA-ConvNeXt model, personnel identity recognition is carried out through palm print information.

Description

Technical Field

[0001] The invention relates to an application method of a SMA-ConvNeXt model in palmprint information personnel identity recognition. Background Art

[0002] With the widespread popularity of information systems, a large amount of personal information is transmitted through the Internet, and people have put forward higher requirements for the protection of private information, the security of system encryption technology, and the accuracy of identity recognition. Biometric recognition plays an important role in personal identity authentication and has been widely used in many fields such as public security, access control systems, judicial evidence collection, and banking.

[0003] Relevant research shows that palmprint recognition, as an important method of biometric identification, is being widely used in various fields. With its outstanding performance in various application scenarios, palmprint recognition has attracted more and more attention. Compared with face, iris, fingerprint and ear recognition, palmprint recognition has significant advantages in dealing with angle changes, lighting effects, collection difficulty and counterfeit protection. Although the existing deep learning methods have improved the performance of palmprint recognition to a certain extent, there are still some challenges, such as the diversity of palmprint features, changes in palm postures, and loss of details under complex lighting conditions. Summary of the invention

[0004] The present invention aims to solve the problems existing in the above-mentioned prior art and provides an application method based on the SMA-ConvNeXt model in the identification of personnel using palmprint information.

[0005] The technical solutions adopted in the present invention are:

[0006] An application method based on the SMA-ConvNeXt model in palmprint information personnel identity recognition includes the following steps:

[0007] S1: Create the SMPDRED infrared integrated palmprint information recognition benchmark test set, and divide the data set into a training set and a validation set after preprocessing the data set;

[0008] S2: Build a ConvNeXt basic model, initialize the model parameters, and train the basic model using a training set;

[0009] S3: An improved SMA-ConvNeXt model obtained by improving the ConvNeXt base model, wherein the improvements include:

[0010] Improvement 1: Propose a gated self-attention mechanism in the ConvNeXt basic model;

[0011] Improvement 2: Propose an adaptive channel enhancement module in the ConvNeXt basic model;

[0012] Improvement 3: Propose a multi-spectral mixed convolution kernel encoder in the ConvNeXt basic model;

[0013] S4: Verify the improved SMA-ConvNeXt model and compare the model indicators with the ConvNeXt base model;

[0014] S5: Personnel identity recognition through palmprint information based on the improved SMA-ConvNeXt model.

[0015] Furthermore, a gated self-attention mechanism is proposed in the ConvNeXt basic model, including: introducing relative position bias in the traditional self-attention mechanism, changing the traditional masking method to sparse masking, and adjusting the output through a gating mechanism.

[0016] Furthermore, the processing flow of feature maps in improvement point 1 is as follows:

[0017] 1) Input feature map:

[0018] Input feature map F, whose shape is C×H×W, generates query vector Q, key vector K and value vector V through linear transformation;

[0019] 2) Calculate the relative position deviation:

[0020] Generate a relative position deviation R for each pair of feature positions (i, j) ij :

[0021] R ij =Embedding(ji)

[0022] R ij It is the relative position deviation information generated by learnable embedding, which is used to capture the position information relationship in the feature map F;

[0023] 3) Calculate the sparse mask:

[0024] Generate a sparse mask matrix M ij , keeping only the important attention calculations:

[0025]

[0026] Invalid attention scores are set to negative infinity to be ignored in softmax;

[0027] 4) Calculate the sparse attention score:

[0028] Introducing relative position deviation R ij and the sparse mask matrix M ij , adjust the attention score calculation formula:

[0029]

[0030] Q i The vector representing the i-th position of the query vector (Q), K j The vector representing the jth position of the key vector (K), d k represents the dimension of the key vector (K);

[0031] 5) Calculate attention weight:

[0032] Perform a softmax operation on the sparse attention scores:

[0033] A ij =softmax(Attention Score ij )

[0034] 6) Use the gate vector g to dynamically adjust the attention output:

[0035] g=σ(W g ·x+b g )

[0036] Attention gated =g⊙(A·V)

[0037] W g is the weight matrix parameter, x is the input feature map F, σ is the Sigmoid activation function, b g is the bias term, ⊙ is the element-by-element multiplication, and V is the value vector.

[0038] Furthermore, an adaptive channel enhancement module is proposed in the ConvNeXt basic model, including: implementing dynamic channel enhancement through adaptive complementary pooling, channel reorganization and channel shuffling operations in the traditional channel enhancement module.

[0039] Furthermore, the processing flow of feature maps in improvement point 2 is as follows:

[0040] 1) Input feature map:

[0041] Input feature map F, shape is C×H×W;

[0042] 2) Global pooling:

[0043] Perform global average pooling on each channel to extract global context information, use the generated channel weights to adaptively enhance the input feature map F, and extract the key features of the overall texture information in the palmprint image. The global feature statistical function is:

[0044]

[0045] Among them, c is the feature map channel index, i is the pixel coordinate in the height direction of the feature map, and j is the pixel coordinate in the width direction of the feature map;

[0046] 3) Local pooling:

[0047] Use a sliding window to calculate local features block by block on the feature map, extract local texture features of the palmprint image, and retain the detailed information of the palmprint. The local texture feature extraction function is:

[0048]

[0049] Among them, k h is the height of local pooling, c is the current channel, k w is the width of local pooling, u and v are the center offsets;

[0050] 4) Dynamically integrate global and local features:

[0051] Use adaptive weights α and β to weightedly fuse global and local features, and the adaptive enhancement function is:

[0052] F pool =α·F global +β·F local

[0053] F global is the global feature, F local It is a local feature;

[0054] α, β are dynamically generated through a learnable network:

[0055] [α,β]=Softmax(MLP(F))

[0056] MLP is a fully connected network used for nonlinear transformation and feature mapping, and F is the input feature map;

[0057] 5) Channel reorganization and shuffling:

[0058] The fused features are reorganized and shuffled to enhance the interactivity between channels. The function is:

[0059] F ACE =ChannelShuffle(Reorganize(F pool ))

[0060] Reorganize represents a channel reorganization operation, and ChannelShuffle represents a channel shuffle operation.

[0061] Furthermore, a multi-spectral mixed convolution kernel encoder is proposed in the ConvNeXt basic model, including replacing the 1×1 convolution in the traditional encoder with a mixed convolution structure of different sizes, and combining the residual block with the global channel attention module to enhance the fusion and modeling capabilities of multi-spectral and multi-scale features.

[0062] Furthermore, the processing flow of feature maps in improvement point 3 is as follows:

[0063] 1) Input feature map:

[0064] Input feature map F, shape is C×H×W;

[0065] 2) Multi-scale convolution kernel feature extraction:

[0066] Use convolution kernels of different sizes, 3×3 and 5×5, to extract multi-scale features from palmprint images:

[0067] F multi =W 3×3 *F+W 5×5 *F+W 3×3 *F

[0068] Among them, W 3×3 ,W 5×5 ,W 3×3 are the weights corresponding to different convolution kernels, and * represents the convolution operation;

[0069] 3) Introducing residual block enhancement features:

[0070] Use residual blocks to enhance the multi-scale features of palmprints:

[0071] F res =F multi +ResBlock(F multi );

[0072] ResBlock is the residual operation:

[0073] ResBlock(x)=Conv2(GELU(Conv1(x)))

[0074] x is the input feature, Conv1 is the convolution operation applied to the input x, GELU is the activation function operation, and Conv2 is the second convolution operation;

[0075] 4) Global channel attention optimization features:

[0076] Use global average pooling to generate the weight of each channel and extract global palmprint information:

[0077]

[0078] Among them, c is the feature map channel index, i is the pixel coordinate in the height direction of the feature map, and j is the pixel coordinate in the width direction of the feature map;

[0079] Adjust channel weights through the global channel attention module:

[0080] F output (c,i,j)=w c ·F res (c,i,j)

[0081] w c represents the weighting coefficient of the cth channel, F res Represents the value of the input residual feature map (Residual Feature) at the cth channel, ith row, and jth column.

[0082] The present invention has the following beneficial effects:

[0083] The present invention uses the SMA-ConvNeXt model to identify the identity of a person in palmprint information. In view of the problem of insufficient diversity in the data set, a self-made SMPDRED infrared integrated palmprint information recognition benchmark test set is proposed. In view of the feature offset and misalignment problems caused by different palm postures, a gated self-attention mechanism is designed to introduce relative position deviation, change the traditional masking method to a sparse mask, and adjust the output through a gating mechanism to improve the global feature expression ability of the model. In addition, in view of the channel feature collapse problem caused by diversified palmprint features, an adaptive channel enhancement module is designed to enhance the information content of key channels and dynamically adjust the weight of each channel. In order to cope with the problem of loss or blurring of detail information caused by complex lighting conditions, a multi-spectral mixed convolution kernel encoder is designed to enrich the diversity and discrimination ability of palmprint features by combining spectral information of different wavelengths and different scale features. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a flow chart of the present invention.

[0085] Figure 2 This is the algorithm structure diagram of the SMA-ConvNeXt model.

[0086] Figure 3 This is the basic structure diagram of the gated self-attention mechanism.

[0087] Figure 4 This is the basic structure diagram of the adaptive channel enhancement module.

[0088] Figure 5 This is the basic structure diagram of the encoder.

[0089] Figure 6 This is the basic structure diagram of the decoder.

[0090] Figure 7 This is the accuracy curve. DETAILED DESCRIPTION

[0091] The present invention will be further described below in conjunction with the accompanying drawings.

[0092] like Figure 1 The present invention discloses an application method of the SMA-ConvNeXt model in palmprint information personnel identification, comprising the following steps:

[0093] S1: Create the SMPDRED infrared integrated palmprint information recognition benchmark test set, and divide the data set into a training set and a validation set after preprocessing the data set;

[0094] S2: Build a ConvNeXt basic model, initialize the model parameters, and train the basic model using a training set;

[0095] S3: An improved SMA-ConvNeXt model is obtained by improving the ConvNeXt basic model, wherein the improvements are as follows:

[0096] (1) Propose a gated self-attention mechanism in the ConvNeXt basic model:

[0097] Relative position bias is introduced into the traditional self-attention mechanism, the traditional masking method is changed to sparse masking, and the output is adjusted through the gating mechanism.

[0098] (2) Propose an adaptive channel enhancement module in the ConvNeXt basic model:

[0099] In the traditional channel enhancement module, adaptive complementary pooling, channel reorganization and channel shuffling operations are used to effectively enhance the information expression ability of key channels, while achieving the optimization of dynamic channel enhancement.

[0100] (3) A multi-spectral mixed convolution kernel encoder is proposed in the ConvNeXt basic model:

[0101] In the traditional encoder, the 1×1 convolution is replaced by a mixed convolution structure of different sizes, and the residual block is combined with the global channel attention module to enhance the fusion and modeling capabilities of multi-spectral and multi-scale features.

[0102] S4: Verify the improved SMA-ConvNeXt model and compare the model indicators with the ConvNeXt base model;

[0103] S5: Personnel identity recognition through palmprint information based on the improved SMA-ConvNeXt model.

[0104] The present invention uses the SMA-ConvNeXt model to identify the identity of a person in palmprint information. In view of the problem of insufficient diversity in the data set, a self-made SMPDRED infrared integrated palmprint information recognition benchmark test set is proposed. In view of the feature offset and misalignment problems caused by different palm postures, a gated self-attention mechanism is designed to introduce relative position deviation, change the traditional masking method to a sparse mask, and adjust the output through a gating mechanism to improve the global feature expression ability of the model. In addition, in view of the channel feature collapse problem caused by diversified palmprint features, an adaptive channel enhancement module is designed to enhance the information content of key channels and dynamically adjust the weight of each channel. In order to cope with the problem of loss or blurring of detail information caused by complex lighting conditions, a multi-spectral mixed convolution kernel encoder is designed to enrich the diversity and discrimination ability of palmprint features by combining spectral information of different wavelengths and different scale features.

[0105] The present invention adopts the self-made SMPDRED infrared integrated palmprint information recognition benchmark test set, which is composed of some public data sets SMPD visible light palmprint images and CASIA multispectral palmprint images.

[0106] The SMPD palmprint dataset mainly includes samples collected under different lighting environments, including diverse samples of different genders, age groups, and postures. The CASIA multispectral dataset mainly contains multispectral images, which can clearly capture the palmprint texture and deep vascular information. 40 representative images of each category are randomly selected to form a new dataset. Random cropping and random horizontal flipping are then used to increase data diversity and reduce the model's dependence on specific image sizes and directions. Secondly, the image is converted into a tensor and standardized using a predefined mean and standard deviation so that the pixel values ​​of each channel conform to the standard normal distribution, thereby accelerating the convergence of the model and improving its performance.

[0107] Figure 2 This is the algorithm structure diagram of the improved SMA-ConvNeXt model. Figure 3 This is the basic structure diagram of the gated self-attention mechanism, which mainly generates relative position deviations through a learnable embedding function. The embedding function dynamically models the relative relationship between features at different positions in the palmprint image, thereby enhancing the local and global correlation of the palmprint; the sparse mask is dynamically generated through the Top-k sparsification strategy, retaining only the k positions with the highest attention scores, effectively reducing the computational complexity, and the gating mechanism dynamically adjusts the attention output, generates gating weights through the Sigmoid activation function, improves the model's ability to allocate the importance of different palmprint features, and finally outputs the optimized attention features. The specific process is:

[0108] 1) Input feature map:

[0109] The input feature map F, whose shape is C×H×W, is linearly transformed to generate the query vector Q, key vector K and value vector V.

[0110] 2) Calculate the relative position deviation:

[0111] Generate a relative position deviation R for each pair of feature positions (i, j) ij :

[0112] R ij =Embedding(ji)

[0113] R ij The relative position deviation information generated by learnable embedding can capture the position information relationship in the feature, thereby enhancing the local and global correlation of the palmprint.

[0114] 3) Calculate the sparse mask:

[0115] Generate a sparse mask matrix M ij , keeping only the important attention calculations:

[0116]

[0117] Invalid attention scores are set to negative infinity so that they can be ignored in the softmax, effectively reducing the computational complexity.

[0118] 4) Calculate the sparse attention score:

[0119] Introducing relative position deviation R ij and the sparse mask matrix M ij , adjust the attention score calculation formula:

[0120]

[0121] Q i The vector representing the i-th position of the query vector (Q), K j The vector representing the jth position of the key vector (K), d k represents the dimension of the key vector (K);

[0122] 5) Calculate attention weight:

[0123] Perform a softmax operation on the sparse attention scores:

[0124] A ij =softmax(Attention Score ij )

[0125] 6) Use the gate vector g to dynamically adjust the attention output:

[0126] g=σ(W g ·x+b g )

[0127] Attention gated =g⊙(A·V)

[0128] W g is the weight matrix parameter, x is the input feature F, σ is the Sigmoid activation function, b g is the bias term, ⊙ is the element-by-element multiplication, and V is the value vector. This effectively improves the model's ability to allocate the importance of different palmprint features, and finally outputs the optimized attention features.

[0129] like Figure 4 Shown is the basic structure diagram of the adaptive channel enhancement module.

[0130] The global and local features are mainly extracted through adaptive complementary pooling. Global pooling is used to extract the overall texture information of the palmprint image, thereby enhancing the model's perception of the global context; local pooling is used to extract the local texture features of the palmprint image and effectively retain the detailed information of the palmprint. Adaptive weights are used to dynamically fuse global and local features to achieve optimal integration of features. In addition, the feature distribution between different channels of the palmprint image is dynamically adjusted through the channel reorganization operation, and the channel shuffling operation rearranges the channel order to enhance the information interaction between channels and further improve the expression ability of palmprint features. The specific process is:

[0131] 1) Input feature map:

[0132] Input feature map F, the shape is C×H×W.

[0133] 2) Global pooling:

[0134] Perform global average pooling on each channel to extract global context information, use the generated channel weights to adaptively enhance the input feature map, extract the key features of the overall texture information in the palmprint image, and enhance the model's ability to perceive the global context. The global feature statistical function is:

[0135]

[0136] Among them, c is the feature map channel index, i is the pixel coordinate in the height (row) direction of the feature map, and j is the pixel coordinate in the width (column) direction of the feature map.

[0137] 3) Local pooling:

[0138] Use a sliding window to calculate local features block by block on the feature map, extract local texture features of the palmprint image, and retain the detailed information of the palmprint. The local texture feature extraction function is:

[0139]

[0140] Among them, k h is the height of local pooling, and c is the current channel. k w is the width of local pooling. u,v are the center offsets.

[0141] 4) Dynamically integrate global and local features:

[0142] Use adaptive weights α and β to weightedly fuse global and local features, and the adaptive enhancement function is:

[0143] F pool =α·F global +β·F local

[0144] F global is the global feature, F local It is a local feature;

[0145] α, β are dynamically generated through a learnable network:

[0146] [α,β]=Softmax(MLP(F))

[0147] MLP is a fully connected network used for nonlinear transformation and feature mapping, and F is the input feature map;

[0148] 5) Channel reorganization and shuffling:

[0149] The fused features are reorganized and shuffled to enhance the interactivity between channels and improve the expression ability of palmprint features. The function is:

[0150] F ACE =ChannelShuffle(Reorganize(F pool ))

[0151] Reorganize represents a channel reorganization operation, and ChannelShuffle represents a channel shuffle operation.

[0152] Figure 5-6The basic structure diagram of the multi-spectral mixed convolution kernel encoder. In the traditional encoder, the 1×1 convolution is replaced by a mixed convolution structure of different sizes of 3×3 and 5×5 to extract multi-scale features in the palmprint image. The residual block is introduced to adopt a two-layer convolution structure, and feature superposition is achieved through jump connections to avoid gradient disappearance and enhance feature expression capabilities. Global average pooling is performed on the input feature map to extract global palmprint information; channel weights are generated through channel attention weight learning, and the weights are used to perform element-by-element multiplication of input features; the feature fusion and modeling capabilities of multi-spectral palmprint images under different spectra are enhanced to capture the commonalities and differences of palmprints under visible light and infrared light. The reconstruction error between the input image and the reconstructed image is used as the training target; the loss function uses the mean square error (MSE) to minimize the reconstruction error of the model and ensure the accuracy and consistency of palmprint feature representation.

[0153] The specific process is:

[0154] 1) Input feature map:

[0155] Input feature map F, the shape is C×H×W.

[0156] 2) Multi-scale convolution kernel feature extraction:

[0157] Use convolution kernels of different sizes (3×3, 5×5) to extract multi-scale features in palmprint images:

[0158] F multi =W 3×3 *F+W 5×5 *F+W 3×3 *F

[0159] Among them, W 3×3 ,W 5×5 ,W 3×3 are the weights corresponding to different convolution kernels, Represents a convolution operation.

[0160] 3) Introducing residual block enhancement features:

[0161] Use residual blocks to enhance the multi-scale features of palmprints to avoid gradient disappearance:

[0162] F res =F multi +ResBlock(F multi )

[0163] ResBlock is the residual operation:

[0164] ResBlock(x)=Conv2(GELU(Conv1(x)))

[0165] x is the input feature, Conv1 is the convolution operation applied to the input x, GELU is the activation function operation, and Conv2 is the second convolution operation;

[0166] 4) Global channel attention optimization features:

[0167] Use global average pooling to generate the weight of each channel and extract global palmprint information:

[0168]

[0169] Where c is the feature map channel index, i is the pixel coordinate in the height (row) direction of the feature map, and j is the pixel coordinate in the width (column) direction of the feature map.

[0170] The global channel attention module is used to adjust the channel weights, enhance the feature fusion and modeling capabilities of multispectral palmprint images under different spectra, and capture the commonalities and differences of palmprints under visible light and infrared light:

[0171] F output (c,i,j)=w c ·F res (c,i,j)

[0172] w c represents the weighting coefficient of the cth channel, F res Represents the value of the input residual feature map (Residual Feature) at the cth channel, ith row, and jth column.

[0173] The training goal of the encoder is to minimize the input image x and the reconstructed image The reconstruction error between the two is taken into account, and the mean square error (MSE) is used as the loss function to minimize the reconstruction error of the model and ensure the accuracy and consistency of the palmprint feature representation:

[0174]

[0175] Where N is the number of pixels in the image, x i and are the i-th pixel values ​​of the input image and the reconstructed image respectively.

[0176] Figure 6Table 1 is the accuracy curve of the SMA-ConvNeXt model on the self-made SMPDRED infrared integrated palmprint information recognition benchmark test set, and Table 2 is the recognition accuracy results. The accuracy of the SMA-ConvNeXt model in palmprint information personnel identification reached 99.3%. Compared with the ConvNeXt benchmark model, the accuracy of the SMA-ConvNeXt model increased by 2.4%, and the equal error rate (EER) was also reduced accordingly. This further verifies the innovation and effectiveness of using the SMA-ConvNeXt model to identify palmprint information to determine personnel identity.

[0177]

[0178] Table 1

[0179] The above description is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be regarded as within the protection scope of the present invention.

Claims

1. An application method based on SMA-ConvNeXt model in palmprint information personnel identity recognition, characterized by: The steps include: S1: Create the SMPDRED infrared integrated palmprint information recognition benchmark test set, and divide the data set into a training set and a validation set after preprocessing the data set; S2: Build a ConvNeXt basic model, initialize the model parameters, and train the basic model using a training set; S3: An improved SMA-ConvNeXt model obtained by improving the ConvNeXt base model, wherein the improvements include: Improvement 1: Propose a gated self-attention mechanism in the ConvNeXt basic model; Improvement 2: Propose an adaptive channel enhancement module in the ConvNeXt basic model; Improvement 3: Propose a multi-spectral mixed convolution kernel encoder in the ConvNeXt basic model; S4: Verify the improved SMA-ConvNeXt model and compare the model indicators with the ConvNeXt base model; S5: Personnel identity recognition through palmprint information based on the improved SMA-ConvNeXt model.

2. The method for applying the SMA-ConvNeXt model in palmprint information personnel identification as claimed in claim 1, characterized in that: The gated self-attention mechanism proposed in the ConvNeXt basic model includes: introducing relative position bias in the traditional self-attention mechanism, changing the traditional masking method to sparse masking, and adjusting the output through the gating mechanism.

3. The method for applying the SMA-ConvNeXt model in palmprint information personnel identification as claimed in claim 2, characterized in that: Improvement point 1: The processing flow of feature maps is as follows: 1) Input feature map: Input feature map F, whose shape is C×H×W, generates query vector Q, key vector K and value vector V through linear transformation; 2) Calculate the relative position deviation: Generate a relative position deviation R for each pair of feature positions (i, j) ij : R ij =Embedding(j-i) R ij It is the relative position deviation information generated by learnable embedding, which is used to capture the position information relationship in the feature map F; 3) Calculate the sparse mask: Generate a sparse mask matrix M ij , keeping only the important attention calculations: Invalid attention scores are set to negative infinity to be ignored in softmax; 4) Calculate the sparse attention score: Introducing relative position deviation R ij and the sparse mask matrix M ij , adjust the attention score calculation formula: Q i The vector representing the i-th position of the query vector (Q), K j The vector representing the jth position of the key vector (K), d k represents the dimension of the key vector (K); 5) Calculate attention weight: Perform a softmax operation on the sparse attention scores: A ij =softmax(Attention Score ij ) 6) Use the gate vector g to dynamically adjust the attention output: g=σ(W g ·x+b g ) Attention gated =g⊙(A·V) W g is the weight matrix parameter, x is the input feature map F, σ is the Sigmoid activation function, b g is the bias term, ⊙ is the element-by-element multiplication, and V is the value vector.

4. The method for applying the SMA-ConvNeXt model in palmprint information personnel identification as claimed in claim 1, characterized in that: The adaptive channel enhancement module proposed in the ConvNeXt basic model includes: dynamic channel enhancement is achieved through adaptive complementary pooling, channel reorganization and channel shuffling operations in the traditional channel enhancement module.

5. The method for applying the SMA-ConvNeXt model in palmprint information personnel identification as claimed in claim 4, characterized in that: Improvement point 2: The processing flow of feature maps is as follows: 1) Input feature map: Input feature map F, shape is C×H×W; 2) Global pooling: Perform global average pooling on each channel to extract global context information, use the generated channel weights to adaptively enhance the input feature map F, and extract the key features of the overall texture information in the palmprint image. The global feature statistical function is: Among them, c is the feature map channel index, i is the pixel coordinate in the height direction of the feature map, and j is the pixel coordinate in the width direction of the feature map; 3) Local pooling: Use a sliding window to calculate local features block by block on the feature map, extract local texture features of the palmprint image, and retain the detailed information of the palmprint. The local texture feature extraction function is: Among them, k h is the height of local pooling, c is the current channel, k w is the width of local pooling, u and v are the center offsets; 4) Dynamically integrate global and local features: Use adaptive weights α and β to weightedly fuse global and local features, and the adaptive enhancement function is: F pool =α·F global +β·F local F global is the global feature, F local It is a local feature; α, β are dynamically generated through a learnable network: [α,β]=Softmax(MLP(F)) MLP is a fully connected network used for nonlinear transformation and feature mapping, and F is the input feature map; 5) Channel reorganization and shuffling: The fused features are reorganized and shuffled to enhance the interactivity between channels. The function is: F ACE =ChannelShuffle(Reorganize(F pool )) Reorganize represents a channel reorganization operation, and ChannelShuffle represents a channel shuffle operation.

6. The method for applying the SMA-ConvNeXt model in palmprint information personnel identification as claimed in claim 1, characterized in that: A multi-spectral mixed convolution kernel encoder is proposed in the ConvNeXt basic model, which includes replacing the 1×1 convolution in the traditional encoder with a mixed convolution structure of different sizes, and combining the residual block with the global channel attention module to enhance the fusion and modeling capabilities of multi-spectral and multi-scale features.

7. The method for applying the SMA-ConvNeXt model in palmprint information personnel identification as claimed in claim 6, characterized in that: Improvement point 3: The processing flow of feature maps is as follows: 1) Input feature map: Input feature map F, shape is C×H×W; 2) Multi-scale convolution kernel feature extraction: Use convolution kernels of different sizes, 3×3 and 5×5, to extract multi-scale features from palmprint images: F multi =W 3×3 *F+W 5×5 *F+W 3×3 *F Among them, W 3×3 ,W 5×5 ,W 3×3 are the weights corresponding to different convolution kernels, and * represents the convolution operation; 3) Introducing residual block enhancement features: Use residual blocks to enhance the multi-scale features of palmprints: F res =F multi +ResBlock(F multi ); ResBlock is the residual operation: ResBlock(x)=Conv2(GELU(Conv1(x))) x is the input feature, Conv1 is the convolution operation applied to the input x, GELU is the activation function operation, and Conv2 is the second convolution operation; 4) Global channel attention optimization features: Use global average pooling to generate the weight of each channel and extract global palmprint information: Among them, c is the feature map channel index, i is the pixel coordinate in the height direction of the feature map, and j is the pixel coordinate in the width direction of the feature map; Adjust channel weights through the global channel attention module: F output (c,i,j)=w c ·F res (c,i,j) w c represents the weighting coefficient of the cth channel, F res Represents the value of the input residual feature map at the cth channel, ith row, and jth column.

Images