Palm print identification method based on multi-feature fusion

Through the multi-feature fusion method, multi-scale features are extracted using Gabor filtering and GS block module, and channel attention adjustment is combined with SE module and Softmax, which solves the problem of insufficient fusion of multi-scale features in the existing technology, and improves the accuracy and robustness of palm print recognition.

CN120164238APending Publication Date: 2025-06-17INTELLIGENT MFG INST OF HFUT
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510202082.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-02-24
Publication Date
2025-06-17

AI Technical Summary

Technical Problem

The existing palm pattern recognition technology lacks an effective multi-scale feature fusion strategy, which leads to limited performance of deep learning methods in the use of palm pattern image information, and insufficient accuracy and robustness of the recognition results.

Method used

Multi-feature fusion method is adopted to extract multi-scale features through Gabor filtering and GS block module, combine SE module and Softmax to perform channel attention mechanism, dynamically adjust the importance of features of different scales, and use ArcFace loss function to calculate the similarity and generate target feature vectors.

Benefits of technology

It improves the accuracy and robustness of palm print recognition, effectively captures rich texture features, and enhances recognition performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120164238A_ABST
    Figure CN120164238A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of palmprint recognition, and discloses a palmprint recognition method based on multi-feature fusion, and the method comprises the steps: collecting a palmprint image, carrying out the preprocessing of the palmprint image, and generating a palmprint ROI image; extracting a multi-scale feature vector of the palmprint ROI image and a histogram feature vector of the palmprint image; splicing the multi-scale feature vector and the histogram feature vector to generate a target feature vector; and calculating the similarity between the target feature vector and the feature vector of the palmprint in a preset database, and outputting a palmprint recognition result according to the similarity. A hierarchical multi-scale filtering mechanism is utilized to process features of different scales layer by layer, and rich texture features are effectively captured; the importance of different scale features is dynamically adjusted by using a multi-scale channel attention mechanism, so that the network pays more attention to feature information beneficial to a recognition task; the histogram features are extracted and fused with the multi-scale features, so that the recognition result is more effective, the accuracy of the recognition result is improved, and the robustness of the recognition result is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the technical field of palmprint recognition, for example, to a palmprint recognition method based on multi-feature fusion. Background Art

[0002] In recent years, as an important branch in the field of biometric recognition, palmprint recognition has received extensive attention and research. Palmprint recognition performs identity verification by analyzing and comparing the unique features of an individual's palmprint, and has the following significant advantages: (1) Palmprint features are rich and stable, ensuring high-precision recognition; (2) It is easy to operate and has a high user acceptance rate; (3) Palmprints can be collected in a non-contact manner, improving the hygiene of the system; (4) Usually, palmprint collection requires the active cooperation of users, and the palmprint recognition system has good privacy protection. Therefore, palmprint recognition technology is widely used in fields such as security monitoring, access control systems, and financial transactions.

[0003] Existing palmprint recognition can be roughly divided into traditional methods and deep learning methods. Among them, traditional methods mainly include methods based on structural features, methods based on subspace learning, methods based on texture features, and methods based on orientation coding, etc. Feature extraction is performed using manually designed filters and requires prior knowledge. Therefore, there are problems such as difficult parameter adjustment and insufficient generality, and it may not be able to well adapt to changes in different databases or real scenarios, resulting in a decline in recognition performance. Deep learning methods can perform automatic feature extraction and have high robustness. However, existing deep learning methods lack an effective multi-scale feature fusion strategy and cannot fully integrate the insights into the rich texture features of palmprint images in traditional methods into the deep learning network, resulting in limitations in the performance of existing deep learning methods in collecting and utilizing the rich information of palmprint images, and making the accuracy and robustness of palmprint recognition results low.

[0004] It should be noted that the information disclosed in the above background art section is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0005] To have a basic understanding of some aspects of the disclosed embodiments, a simple summary is given below. The summary is not a general review, nor is it intended to identify key / important constituent elements or delineate the protection scope of these embodiments, but rather serves as a preamble to the subsequent detailed description.

[0006] Embodiments of the present disclosure provide a palmprint recognition method based on multi-feature fusion to improve the accuracy of palmprint recognition results and enhance the robustness of palmprint recognition results.

[0007] In one embodiment, the palmprint recognition method based on multi-feature fusion includes: collecting a palmprint image and preprocessing the palmprint image to generate a palmprint ROI image; extracting the multi-scale feature vector of the palmprint ROI image and the histogram feature vector of the palmprint image; splicing the multi-scale feature vector and the histogram feature vector to generate a target feature vector; calculating the similarity between the target feature vector and the feature vector of the palmprint in the preset database, and outputting the palmprint recognition result according to the similarity.

[0008] In one embodiment, extracting the multi-scale feature vector of the palmprint ROI image includes: performing maximum-scale Gabor filtering on the palmprint ROI image to generate a feature map X1; performing an element-wise addition operation on the palmprint ROI image and the feature map X1 to generate a fused feature X'; performing multiple Gabor filtering operations with different smaller scales on the fused feature X' to generate feature maps X2, …, X n ; n is a natural number greater than 2; generating multiple features F1, F2, ……, F of different scales according to the feature maps X1, X2, …, X of different scales n ; splicing the multiple features of different scales and extracting the feature vector to generate a multi-scale feature vector. n

[0009] In one embodiment, the GS block is used to perform Gabor filtering on the palmprint ROI image; wherein, the GS block includes an LGC layer and a Softmax; there are multiple GS blocks, and the scales of each GS block are different.

[0010] In one embodiment, the formula for the LGC layer to extract the Gabor convolution kernel of the feature map is as follows:

[0011]

[0012] where G is the real part of the Gabor filter; J is the center of the real part of the Gabor filter; K is the Gabor kernel; W and H are the width and height of the Gabor kernel; i and j, m and n are pixel indices; i ∈ (0, 1, 2, …, W - 1); j ∈ (0, 1, 2, …, H - 1); the parameters in the LGC layer include learnable parameters Φ = {γ, μ, σ} and fixed parameters θ = {iπ / n, i = 0, 1, …, n}; x' = (x - x0)cosθ + (y - y0)sinθ, y' = (x - x0)sinθ + (y - y0)cosθ; (x0, y0) is the center of the function; θ is the direction of the Gabor function in arc length units; σ is the standard deviation of the Gaussian envelope; μ is the frequency of the sine wave; γ is the spatial aspect ratio.

[0013] In one embodiment, a softmax operation is applied to the feature map Τ passing through the LGC layer, and the formula is as follows:

[0014]

[0015] Among them, in the LGC layer, each channel of the input feature is convolved with filters in n directions to generate response feature maps in n directions, denoted as Τ∈R n×A×B ; d is the channel index; i and j are pixel indices; i∈(0, 1, 2, …, W-1); j∈(0, 1, 2, …, H-1); T′(d, i, j) is the feature map after obtaining the attention weights.

[0016] In one embodiment, multiple features of different scales are concatenated, including: extracting the channel attention of multiple features of different scales using a preset SE module to generate the channel attention weight vector for each scale; using Softmax to recalibrate the features of the channel attention weight vector for each scale to generate the soft assignment weight vector after channel interaction for each scale; weighting and concatenating the features of different scales according to the soft assignment weight vector after channel interaction for each scale.

[0017] In one embodiment, the SE module includes compression and excitation. Extracting the channel attention of multiple features of different scales using a preset SE module includes: in the compression part, using global average pooling operation to generate global feature channel information, and the formula is as follows:

[0018]

[0019] Among them, gp represents pooling all feature maps along the feature channel dimension; F i is the feature map of the i-th scale; H f is the height of the feature map F i ; W f is the width of the feature map F i ;

[0020] In the excitation part, using a fully connected layer to combine the internal linear relationship in the feature channel dimension and obtain the attention weights based on the global features, and the formula is as follows:

[0021] W i =τ(FC1(δ(FC0(gp(F i )))), i = 1, 2, ..., n

[0022] Among them, n is a natural number greater than 2; W i is the channel attention weight vector of the i-th scale; δ is the ReLu activation function; FC0 and FC1 represent fully connected layers; τ is the Sigmoid activation function.

[0023] In one embodiment, the histogram feature vector of the palmprint image is extracted, including: filtering the image with Gabor filters in 6 directions to generate the response value and direction of each pixel; dividing the palmprint image into multiple regions; calculating the histogram feature of each region according to the response value and direction of each pixel; and splicing the histogram features of each region to generate the histogram feature vector.

[0024] In one embodiment, the similarity between the target feature vector and the feature vector of the palmprint in the preset database is calculated using the ArcFace loss function, and the specific formula is as follows:

[0025]

[0026] where N is the batch size, s is the scale parameter, m is the angular margin parameter, y i is the true category of sample i, θ yi is the angle between sample i and its true category, and θ j is the angle between sample i and other categories.

[0027] The palmprint recognition method based on multi-feature fusion provided by the embodiments of the present disclosure can achieve the following technical effects:

[0028] The GS block module is used to extract multi-scale features of the ROI image at multiple levels, so as to process and fuse features of different scales layer by layer through a hierarchical multi-scale filtering mechanism that combines traditional methods and convolutional neural network technologies, gradually refining and enhancing the multi-scale texture features of the palmprint image, and effectively capturing rich texture features. The multi-scale channel attention mechanism that cooperates with the SE module and Softmax dynamically adjusts the importance of features at different scales, generates the weights corresponding to the features, enables the network to pay more attention to the feature information beneficial to the recognition task, and improves the palmprint recognition performance. Extracting the histogram features in the palm picture and fusing them with the multi-scale features makes the recognition result more effective. Furthermore, it helps to improve the accuracy of the recognition result and enhance the robustness of the recognition result.

[0029] The above general description and the following description are only exemplary and explanatory, and are not used to limit this application. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] One or more embodiments are exemplarily illustrated by the corresponding drawings. These exemplary illustrations and the drawings do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are shown as similar elements. The drawings do not constitute a scale limitation, and among them:

[0031] Figure 1 is a schematic diagram of the system module of a multi-scale feature extraction system provided by the embodiments of the present disclosure.

[0032] Figure 2 It is a schematic diagram of a hierarchical multi-scale filtering mechanism based on the Gabor filtering unit GS block provided by an embodiment of the present disclosure;

[0033] Figure 3 It is a schematic diagram of a palmprint recognition method based on multi-feature fusion provided by an embodiment of the present disclosure;

[0034] Figure 4 It is a schematic diagram of a multi-scale channel attention mechanism based on the SE module provided by an embodiment of the present disclosure. Detailed implementation manners

[0035] In order to be able to understand the features and technical content of the embodiments of the present disclosure in more detail, the implementation of the embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference and illustration only and are not intended to limit the embodiments of the present disclosure. In the following technical description, for the sake of explanation, numerous details are provided to give a thorough understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other instances, well-known structures and devices may be shown in a simplified manner to simplify the drawings.

[0036] The terms "first", "second", etc. in the specification and claims of the embodiments of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so as to implement the embodiments of the present disclosure described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion.

[0037] Unless otherwise specified, the term "plurality" means two or more.

[0038] The term "corresponding" may refer to an association relationship or a binding relationship. That A corresponds to B means that there is an association relationship or a binding relationship between A and B.

[0039] It should be noted that, without conflict, the embodiments and the features in the embodiments of the present disclosure may be combined with each other.

[0040] Figure 1 Shows a multi-scale feature extraction system for hierarchically extracting multi-scale features. In combination with Figure 1 As shown, the system environment includes a Gabor filtering unit GS block, a subsequent processing unit SPU, and a feature vector extraction unit FVE.

[0041] Figure 2Shows a schematic diagram of a hierarchical multi-scale filtering mechanism based on the Gabor filtering unit GS block. There are multiple GS blocks, and each GS block has a different scale, which are respectively used to extract features of different scales. Specifically, the GS block includes an LGC layer and a Softmax. The formula for the LGC layer to extract the Gabor convolution kernel of the feature map is as follows:

[0042]

[0043] Where G is the real part of the Gabor filter; J is the center of the real part of the Gabor filter; K is the Gabor kernel; W and H are the width and height of the Gabor kernel; i and j, m and n are pixel indices; i ∈ (0, 1, 2,..., W - 1); j ∈ (0, 1, 2,..., H - 1); the parameters in the LGC layer include learnable parameters Φ = {γ, μ, σ} and fixed parameters θ = {iπ / n, i = 0, 1,..., n}; x' = (x - x0)cosθ + (y - y0)sinθ, y' = (x - x0)sinθ + (y - y0)cosθ; (x0, y0) is the center of the function; θ is the direction of the Gabor function in arc length units; σ is the standard deviation of the Gaussian envelope; μ is the frequency of the sine wave; γ is the spatial aspect ratio.

[0044] Apply the softmax operation on the channel dimension to the feature map Τ after passing through the LGC layer, and the formula is as follows:

[0045]

[0046] Where, in the LGC layer, each channel of the input feature is convolved with filters in n directions to generate n response feature maps in different directions, denoted as Τ ∈ R n×A×B ; d is the channel index; i and j are pixel indices; i ∈ (0, 1, 2,..., W - 1); j ∈ (0, 1, 2,..., H - 1); T'(d, i, j) is the feature map after obtaining the attention weights.

[0047] Combined with Figure 1 As shown, the subsequent processing unit SPU includes a max pooling layer and a common convolution layer. Input the feature map X i of the i-th scale into the SPU unit to obtain the feature map F i representing the i-th scale, and the calculation formula is as follows:

[0048]

[0049] Where C is the number of channels of the feature F i ; H f is the height of the feature F i ; W f is the width of the feature F iWidth.

[0050] Combined with Figure 1 and Figure 2 As shown, in this embodiment, the scale of the GS block is in large scale GS L medium scale GS M and small scale GS S as an example. That is, in this embodiment, the extraction of multi-scale features is illustrated by taking the extraction of features of large, medium, and small scales as an example.

[0051] Combined with Figure 1 and Figure 2 the system shown, the palmprint recognition method based on multi-feature fusion of the present disclosure is as Figure 3 shown, and the method includes:

[0052] S101, collect a palmprint image and preprocess the palmprint image to generate a palmprint ROI image.

[0053] In the embodiment of the present invention, the preprocessing includes steps such as cropping the palmprint image and image grayscale processing, so as to generate a palmprint ROI image.

[0054] S102, extract the multi-scale feature vector of the palmprint ROI image and the histogram feature vector of the palmprint image.

[0055] When extracting the multi-scale feature vector of the palmprint ROI image, it is extracted by using a multi-scale feature extraction system. The multi-scale feature extraction system can utilize the hierarchical multi-scale filtering mechanism of the GS block module to extract the multi-scale features of the ROI image, and then utilize the preset multi-scale channel attention mechanism to extract the channel attention of different scale feature maps, and fuse the multi-scale features according to the channel attention of different scale feature channel maps to generate a multi-scale feature vector.

[0056] When extracting the histogram feature vector of the palmprint image, it is extracted by using a preset Gabor filter in 6 directions.

[0057] S103, splice the multi-scale feature vector and the histogram feature vector to generate a target feature vector.

[0058] S104, calculate the similarity between the target feature vector and the feature vector of the palmprint in the preset database, and output the palmprint recognition result according to the similarity.

[0059] Using a hierarchical multi-scale filtering mechanism, process and fuse features of different scales layer by layer, gradually refine and enhance the multi-scale texture features of the palmprint image, and effectively capture rich texture features; use a multi-scale channel attention mechanism to dynamically adjust the importance of features of different scales, so that the network pays more attention to the feature information beneficial to the recognition task; extract the histogram features in the palm image and fuse them with the multi-scale features to make the recognition result more effective, thereby helping to improve the accuracy of the recognition result and enhance the robustness of the recognition result.

[0060] In one embodiment, extracting a multi-scale feature vector of the palmprint ROI image includes: performing maximum-scale Gabor filtering on the palmprint ROI image to generate a feature map X1; performing an element-wise addition operation on the palmprint ROI image and the feature map X1 to generate a fused feature X'; performing multiple Gabor filtering operations with different smaller scales on the fused feature X' respectively to generate feature maps X2, …, X n ; n is a natural number greater than 2; according to the feature maps X1, X2, …, X of different scales n generate multiple features F1, F2, ……, F of different scales n ; splice the features of multiple different scales and extract the feature vector to generate a multi-scale feature vector.

[0061] Combined Figure 1 and Figure 2 As shown, perform Gabor filtering on the palmprint ROI image using the GS block; among them, there are multiple GS blocks, and the scales of each GS block are different. In this embodiment, the scales of the GS blocks are large-scale GS L , medium-scale GS M and small-scale GS S as an example. That is, in this embodiment, the extraction of multi-scale features is illustrated by taking the extraction of large, medium, and small scale features as an example. Specifically, in this embodiment, in the GS L block, the kernel size in the LGC layer is 11×11, the number is 24, and the stride is 1; in the GS M block, the kernel size of the LGC layer is 7×7, the number is 24, and the stride is 1; in the GS S block, the kernel size of the LGC layer is 3×3, the number is 24, and the stride is 1.

[0062] First, input the ROI image into the GS L block to perform large-scale Gabor filtering, and the output feature map is denoted as X1. Then perform an element-wise addition operation on the ROI image and the feature map X1 to obtain the fused feature X'. Input the fused feature X' in parallel into the GS of a smaller scale M block and GS SIn the block, the feature maps X2 and X3 are obtained respectively. Then, the feature maps X1, X2, and X3 are respectively input into the subsequent processing unit SPU, and the output features are F1, F2, and F3 respectively, and the feature sizes are all 24×30×30. Finally, the features F1, F2, and F3 of different scales are concatenated to obtain the feature F containing different scales, and the feature size is 72×30×30. The specific concatenation formula is:

[0063]

[0064] Among them, is the concatenation operator, and C is the number of channels.

[0065] The GS block module is used to extract the multi-scale features of the ROI image at multiple levels, so as to process and fuse the features of different scales layer by layer through a hierarchical multi-scale filtering mechanism that combines traditional methods and convolutional neural network technologies, gradually refining and enhancing the multi-scale texture features of the palmprint image, and effectively capturing rich texture features.

[0066] In one embodiment, concatenating features of multiple different scales includes: using the SE module to extract the channel attention of the features of multiple different scales to generate the channel attention weight vector at each scale; using Softmax to re-calibrate the features of the channel attention weight vector at each scale to generate the soft assignment weight vector after channel interaction at each scale; weighting and concatenating the features of different scales according to the soft assignment weight vector after channel interaction at each scale.

[0067] Combined Figure 4 As shown, the SE module includes two parts: compression and excitation. In the compression part, the global average pooling operation is used to generate the global feature channel information. gp represents the pooling operation on all feature maps along the feature channel dimension, and the calculation formula is as follows:

[0068]

[0069] Among them, F i represents the feature F i at the i-th scale; H f is the height of the feature F i ; W f is the width of the feature F i .

[0070] In the excitation part, taking the GS blocks of three scales as an example, the fully connected layer is used to combine the internal linear relationship in the feature channel dimension to obtain the attention weight based on the global feature. The channel attention weight vector W i of the i-th scale is calculated as follows:

[0071] W i= τ(FC1(δ(FC0(gp(F i )))), i = 1, 2, ..., n

[0072] where n is a natural number greater than 2; δ represents the ReLu activation function, FC0 and FC1 represent fully connected layers, and τ represents the Sigmoid activation function. Here, when there are three scales, n = 3.

[0073] The multi-scale channel attention vectors are re-calibrated using Softmax to obtain the soft assignment weights W i ′ after the new multi-scale channel interaction, and the re-calibrated attention weights are fused and concatenated to obtain the entire multi-scale attention weight W′. The calculation formula is as follows:

[0074]

[0075] W′ ∈ R 3C×1×1

[0076] where C is the number of channels.

[0077] The channel attention of the feature maps at different scales is extracted using the SE module to obtain the channel attention vectors at each different scale. The SE module consists of two parts: compression and excitation. In the compression part, the global average pooling operation gp is used to generate the global feature channel information; in the excitation part, the attention weight W i is obtained based on the global features. Then, the multi-scale channel attention vectors are re-calibrated using Softmax to obtain the attention weights W i ′ of different channels after the new multi-scale channel interaction, and they are fused and concatenated to obtain the entire multi-scale attention weight W′. The element-wise product operation is performed on the re-calibrated weights and the corresponding feature maps to obtain the calibrated feature F. The feature F is flattened using the feature vector extraction unit, and after feature vector extraction, the feature vector Fe containing more discriminative multi-scale information is obtained. The calculation formula is as follows:

[0078] Fe = FC(D(δ(BN(FC(Flatten(F′)))))

[0079] where FC is the fully connected layer; BN is the 1D batch normalization layer, δ is the ReLu activation function; D is Dropout.

[0080] Based on channel attention, the idea of multi-scale is introduced, enabling the neural network to have stronger representation ability. The multi-scale channel attention mechanism that cooperates with the SE module and Softmax dynamically adjusts the importance of features at different scales, generates the weights corresponding to the features, enables the network to pay more attention to the feature information beneficial to the recognition task, and improves the performance of palmprint recognition.

[0081] In one embodiment, the histogram feature vector of the palmprint image is extracted, including: filtering the image with Gabor filters in 6 directions to generate the response value and direction of each pixel; dividing the palmprint image into multiple regions; calculating the histogram feature of each region according to the response value and direction of each pixel; splicing the histogram features of each region to generate the histogram feature vector.

[0082] In this embodiment, the formula of the Gabor filter is as follows:

[0083]

[0084] where θ is the filter direction and σ is the standard deviation of the Gaussian function.

[0085] In this embodiment, the calculation formulas of the response value m(x, y) and the direction θ(x, y) are as follows:

[0086] m(x,y)=min(I(x,y)*G(x,y,θ k ))

[0087] θ(x,y)=argmin k (I(x,y)*G(x,y,θ k ))

[0088] where k ∈ [1, 6].

[0089] In one embodiment, the similarity between the target feature vector and the feature vector of the palmprint in the preset database is calculated using the ArcFace loss function, and the specific formula is as follows:

[0090]

[0091] where N is the batch size, s is the scale parameter, m is the angular margin parameter, y i is the true category of sample i, θ yi is the angle between sample i and its true category, θ j is the angle between sample i and other categories.

[0092] The above description and the accompanying drawings fully illustrate the embodiments of the present disclosure, enabling those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, process, and other changes. Embodiments merely represent possible variations. Unless explicitly required, individual components and functions are optional, and the order of operations may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terms used in this application are only for describing embodiments and do not limit the claims. As used in the description of embodiments and claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to also include the plural forms. Similarly, as used in this application, the term "and / or" refers to any and all possible combinations including one or more of the associated listed items. Additionally, when used in this application, the term "comprise" and its variants "comprises" and / or "comprising" etc. mean the presence of the stated features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or groups of these. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, or apparatus comprising the element. Herein, each embodiment may focus on the differences from other embodiments, and the same or similar parts among the embodiments may be referred to each other. For the methods, products, etc. disclosed in the embodiments, if they correspond to the method parts disclosed in the embodiments, the relevant parts may refer to the description of the method parts.

[0093] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner may depend on the specific application and design constraints of the technical solution. The skilled person can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the embodiments of the present disclosure. The skilled person can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

Claims

1. A palmprint recognition method based on multi-feature fusion, characterized in that: include: Collecting palmprint images and preprocessing the palmprint images to generate palmprint ROI images; Extracting multi-scale feature vectors of palmprint ROI image and histogram feature vectors of palmprint image; Concatenate the multi-scale feature vector and the histogram feature vector to generate the target feature vector; Calculate the similarity between the target feature vector and the feature vector of the palmprint in the preset database, and output the palmprint recognition result based on the similarity.

2. The palmprint recognition method based on multi-feature fusion according to claim 1 is characterized in that: Extract multi-scale feature vectors of palmprint ROI images, including: Perform maximum-scale Gabor filtering on the palmprint ROI image to generate feature map X1; Perform element-wise addition operation on the palmprint ROI image and feature map X1 to generate fusion feature X ′ ; For the fusion feature X ′ Perform multiple smaller-scale Gabor filtering processes at different scales to generate feature maps X2, ..., X n ; n is a natural number greater than 2; According to the feature maps X1, X2, ..., X n Generate multiple features of different scales F1, F2, ..., F n ; Multiple features of different scales are concatenated and feature vectors are extracted to generate multi-scale feature vectors.

3. The method according to claim 2, characterized in that The GS block is used to perform Gabor filtering on the palmprint ROI image; wherein the GS block includes an LGC layer and a Softmax; there are multiple GS blocks, and each GS block has a different scale.

4. The method according to claim 3, characterized in that The formula of the Gabor convolution kernel used by the LGC layer to extract features is as follows: Where G is the real part of the Gabor filter; J is the real center of the Gabor filter; K is the Gabor kernel; W and H are the width and height of the Gabor kernel; i and j, m and n are pixel indices; the parameters in the LGC layer include learnable parameters Φ = {γ, μ, σ} and fixed parameters θ = {iπ / n, i = 0, 1, ..., n}; x ′ =(x-x0)cosθ+(y-y0)sinθ,y ′ =(x-x0)sinθ+(y-y0)cosθ; (x0, y0) is the center of the function; θ is the direction of the Gabor function in units of arc length; σ is the standard deviation of the Gaussian envelope; μ is the frequency of the sine wave; γ is the spatial aspect ratio.

5. The method according to claim 4, characterized in that The softmax operation on the channel dimension is applied to the feature map Τ after the LGC layer, and the formula is as follows: In the LGC layer, each channel of the input feature is convolved with the filter in n directions to generate a response feature map in n directions, denoted as Τ∈R n×A×B ; d is the channel index; i and j are pixel indices; i∈(0, 1, 2, ..., W-1); j∈(0, 1, 2, ..., H-1); T ′ (d, i, j) is the feature map after obtaining the attention weights.

6. The method according to claim 2, characterized in that Concatenate features from multiple scales, including: Use the preset SE module to extract the channel attention of features at multiple scales and generate the channel attention weight vector at each scale; Use Softmax to recalibrate the channel attention weight vector at each scale to generate a soft assigned weight vector after the channel interaction at each scale; Features of different scales are weighted and concatenated according to the soft assigned weight vector after the interaction of each scale channel.

7. The method according to claim 1, characterized in that The SE module includes compression and excitation. The channel attention of extracting multiple features of different scales using the preset SE module includes: In the compression part, the global average pooling operation is used to generate global feature channel information. The formula is as follows: Among them, gp represents the pooling operation of all feature maps along the feature channel dimension; F i is the feature map of the i-th scale; H f is the feature map F i Height; W f is the feature map F i Width; In the motivation part, the linear relationship within the feature channel dimension combination of the fully connected layer is used to obtain the attention weight based on the global feature. The formula is as follows: W i =τ(FC1(δ(FC0(gp(F i )))),i=1,2,...,n Where n is a natural number greater than 2; W i is the channel attention weight vector of the i-th scale; δ is the ReLu activation function; FC0 and FC1 represent the fully connected layers; τ is the Sigmoid activation function.

8. The method according to claim 1, characterized in that: Extract the histogram feature vector of the palmprint image, including: The image is filtered using Gabor filters in six directions to generate the response value and direction of each pixel; Dividing the palmprint image into multiple regions; Calculate the histogram features of each region based on the response value and direction of each pixel; The histogram features of each region are concatenated to generate a histogram feature vector.

9. The method according to claim 1, characterized in that: The ArcFace loss function is used to calculate the similarity between the target feature vector and the feature vector of the palmprint in the preset database. The specific formula is as follows: Where N is the batch size, s is the scale parameter, m is the angle interval parameter, and y i is the true category of sample i, θ yi is the angle between sample i and its true category, θ j is the angle between sample i and other categories.

Citation Information

Cited By

  • Palm vein and palm print ROI extraction method and system based on multi-feature point fusion

    CN120823627A