A method for high-resolution palmprint recognition based on a multi-input convolutional neural network
By performing adaptive ROI cropping and designing multi-input convolutional neural network model on high-resolution palm print images, the problems of incomplete information and contradiction between decision-making rules in high-resolution palm print recognition are solved, and the recognition accuracy and efficiency are improved.
Patent Information
- Application Number
- CN202210176443.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-25
- Publication Date
- 2025-08-01
- Estimated Expiration
- 2042-02-25
AI Technical Summary
The prior art has problems such as incomplete information and contradiction in decision-making rules in high-resolution palm print recognition, resulting in low recognition accuracy.
Using a method based on multi-input convolutional neural network, a multi-input convolutional neural network model is designed by adaptive ROI cropping of high-resolution palm print images, feature extraction and fusion is used for center and edge ROI features, and chi-square distance calculation and Adam optimizer are used for training.
It improves the accuracy and efficiency of high-resolution palm print recognition, makes full use of palm print feature information, and achieves a more efficient recognition effect.
Smart Images
Figure CN114581962B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of high - resolution palmprint recognition, and particularly to a method for high - resolution palmprint recognition. Background Art
[0002] The world is in the wave of the digital age, and biometric recognition technology has developed rapidly both in theoretical research and practical applications. Biometric recognition technology refers to a technology that uses the inherent physiological or behavioral characteristics of the human body by a computer for personal identity authentication. It has many advantages such as good security, not easy to lose, difficult to counterfeit, and convenient to carry. Common biometric recognition technologies include fingerprint and palmprint recognition, iris recognition, face recognition, voice recognition, etc. With the development of artificial intelligence, biometric recognition technology has become more mature. Palmprint refers to the palm image from the fingertips to the wrist. Compared with fingerprints, it has a larger area and can better show characteristic information. Many of its features can be used for identity recognition, such as main lines, wrinkles, fine textures, ridge endings, bifurcation points, etc. Palmprint acquisition is also a non - invasive and non - contact acquisition method, which is relatively easy for users to accept and has low requirements for equipment. Due to its uniqueness, lifelong invariability, and convenience of acquisition, palmprint recognition has become a popular field in biometrics. As an important part of palmprint recognition technology, high - resolution palmprint recognition has gradually attracted people's interest and attention. High - resolution palmprints can extract more accurate features compared to ordinary low - resolution palmprint images, which is beneficial to improving the recognition accuracy.
[0003] A neural network can be simply described as an information - processing system designed to imitate the structure and function of the human brain, with powerful self - learning ability and the function of quickly finding the optimal solution. When using traditional methods for biometric recognition, sometimes the information source is incomplete and contains false images, and the decision rules are sometimes contradictory and sometimes unruly, which brings great difficulties to traditional recognition methods. However, neural networks can handle these problems well and give reasonable recognition and judgment. This makes it possible to recognize palmprints through neural networks. Summary of the Invention
[0004] The present invention aims to overcome the above - mentioned disadvantages of the prior art and proposes a method for high - resolution palmprint recognition based on a multi - input convolutional neural network.
[0005] In order to make the most of the feature information of high-resolution palmprints, the present invention proposes a method for high-resolution palmprint recognition based on a multi-input convolutional neural network. The effective feature information is obtained from palmprint data through the powerful self-learning ability of the neural network for recognition, eliminating the complicated feature extraction algorithms and palmprint matching processes, thereby achieving high-efficiency palmprint recognition. The multi-input convolutional neural network has three paths, which respectively receive the feature of the central region of one piece of image, the features of two edge regions (the ROI regions of the interdigital area and the hypothenar area), and the feature of one edge region (the ROI region of the thenar area) as the network input, so as to obtain better accuracy than using only one of the inputs and improve the learning effect.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0007] A method for high-resolution palmprint recognition based on a multi-input convolutional neural network, comprising the following steps:
[0008] 1) Define the high-resolution palmprint ROI, including the cropping of the central ROI region and the edge ROI regions (including the thenar ROI region, the hypothenar ROI region, and the interdigital ROI region, hereinafter referred to as ROI1, ROI2, and ROI3). Therefore, a total of 4 pieces of ROI region images can be cropped from one palm data as the data for recognition.
[0009] 2) Design a network model for high-resolution palmprint recognition, use the ROI samples cropped in step 1), and perform data augmentation on them to generate more palmprint ROI data for training.
[0010] 3) Run the multi-input convolutional neural network model and output the result.
[0011] Furthermore, the step 1) includes:
[0012] (11) The acquisition of palmprint ROI region samples adopts an adaptive ROI cropping algorithm. First, extract the coordinates of the two valley points of the palmprint (between the middle finger and the index finger, and between the ring finger and the little finger), and then establish a two-dimensional rectangular coordinate system based on the two valley points. We use the line between the two valley points as the X-axis and establish the Y-axis perpendicular to this line, with the valley point coordinates between the ring finger and the little finger as the origin.
[0013] (12) For the cropping of the central ROI, according to the distance d between the two valley points, we selected (-0.2d, 0.2d) as the upper left corner coordinates of the cropping area, and then intercepted a length of 1.4d along the positive directions of the X and Y coordinate axes respectively; for other edge ROIs, we obtained the corresponding cropping area sizes based on the statistical analysis of the palmprint data. For ROI1, ROI2, and ROI3, we selected (0.7d, 1.2d), (-0.4d, 0.7d), and (-0.25d, 0.1d) in the coordinate system as the upper left corner coordinate positions of the areas to be cropped respectively, and then intercepted lengths of 0.6d, 0.5d, and d along the positive direction of the X coordinate axis respectively, and lengths of 0.6d, 0.9d, and 0.4d along the positive direction of the Y coordinate axis respectively.
[0014] (13) Therefore, the size of the central ROI area is cropped to 1.4d * 1.4d, and the sizes of ROI1, ROI2, and ROI3 are cropped to 0.6d * 0.6d, 0.5d * 0.9d, and d * 0.4d respectively. Then, the image size is normalized. Finally, the size of the central ROI is 512 * 512, and the sizes of ROI1, ROI2, and ROI3 are 512 * 512, 900 * 300, and 900 * 300 respectively. These are used as samples for training the network.
[0015] Furthermore, step (2) includes:
[0016] (21) Design and construct a multi-input convolutional neural network model, set the training parameters and loss function, so that it can simultaneously extract and recognize the features of the central area and edge area of the palmprint image. This network is mainly divided into two parts, namely the feature extraction part and the feature fusion part.
[0017] (211) The first part is the feature extraction part. The feature extraction part performs downsampling operations on the high-resolution palmprint ROI. Due to the problem of the large size difference in a set of palmprint ROI data, the feature extraction part is divided into three paths, and the central ROI, ROI2 and ROI3, and ROI1 are input respectively.
[0018] First, for the path of central ROI image feature extraction, it consists of 7 layers. Each layer is composed of two convolutional modules and one pooling layer. The convolutional kernel size of the convolutional layer in each convolutional module is 3*3, the number is 16, and the stride is 1; the stride of the pooling layer is 2. The input central ROI image passes through the first convolutional layer, BN layer, and ReLU activation function of the first convolutional module in sequence. Then it passes through the first convolutional layer, BN layer, and ReLU activation function of the second convolutional module in sequence, and finally undergoes downsampling processing by the pooling layer. The subsequent processing is the same as above, and convolution followed by pooling is performed 6 more times. The convolutional kernel sizes of the corresponding convolutional modules are all 3*3, and the numbers are 32, 64, 128, 256, 512, and 512 respectively. Input the central ROI image with a size of 512×512, pass through the first layer to the seventh layer in sequence, and the output feature size is 4*4*512;
[0019] For the feature extraction paths of edge ROIs 2 and 3, they are similar to the previous central ROI feature extraction path. However, due to the relatively small image size, one convolution-pooling step is missing, and it consists of 6 layers. The number of convolutional kernels corresponding to each layer is 32, 64, 128, 256, 512, and 512 respectively. Since the sizes of the 4 ROIs vary greatly, to ensure better and more convenient fusion of the extracted features, during the feature extraction process, when inputting the ROI2 and ROI3 images with a size of 900*300, slicing work is performed. The image is sliced into three parts with a size of 300×300, and then input in sequence; Input the 300×300 image, pass through the first layer to the sixth layer in sequence, and the output feature size is 4*4*512;
[0020] For the ROI1 feature extraction path, it is the same as the central ROI feature extraction path. Input the ROI1 image with a size of 512×512, pass through the first layer to the seventh layer in sequence, and the output feature size is 4*4*512;
[0021] (212) The second part is the feature fusion part. That is, the edge ROI features and central ROI features of the palm are fused. During the above feature extraction process, the final output feature sizes of the central ROI, ROI1, and the sliced ROI2 and ROI3 are all 4*4*512, which are kept consistent. Then, the features output by the three paths are cascaded for feature layer fusion. The fused features are then subjected to two convolutional modules and one pooling layer operation. The convolutional kernel size of the convolutional layer in each convolutional module is 3*3, the number is 1024, and the stride is 1; the stride of the pooling layer is 2. Finally, adaptive average pooling is performed to form a 1024-dimensional feature vector.
[0022] (22) The network training part inputs three pieces of data, namely the anchor, the positive sample, and the negative sample. The multi-input convolutional neural network is used to extract features from these three pieces of data, so their parameters are shared. Then, the chi-square distance is used to calculate the distances between the anchor and the positive sample, and between the anchor and the negative sample, and these two distances are fed into the loss function for calculation. After that, the Adam Optimizer is used for optimization iteration, and the optimal solution is obtained after multiple iterations.
[0023] In the above training process, the loss function of the multi-input convolutional neural network is:
[0024] Loss = max(d(a, p) - d(a, n) + margin, 0) (1)
[0025] Where a represents the anchor, p represents the positive sample of the same class as a, n represents the negative sample of a different class from a, margin represents a constant, set to 0.2, and d(a, n) represents the distance between the anchor a and the negative sample n. The goal of minimizing Loss is to make d(a, p) close to 0 and d(a, n) greater than d(a, p) + margin.
[0026] Furthermore, step 3) includes:
[0027] Run the network model, input the test palmprint data, use the trained network model to calculate the chi-square distance between this test data and all the reference palmprint data in the database, find the reference palmprint image with the smallest distance value. If this distance is less than the set threshold of 0.5, it is determined to be the same class; otherwise, there is no class the same as the test palmprint image.
[0028] The working principle of the present invention is: redefine the ROI area of the high-resolution palmprint, then perform adaptive cropping of the ROI area, and input it into the designed multi-input convolutional neural network model for feature extraction and feature fusion, thereby improving the accuracy of high-resolution palmprint recognition.
[0029] The advantages of the present invention are: the redefinition and cropping of the high-resolution palmprint ROI can make more full use of the palmprint feature information; the designed multi-input convolutional neural network model can simultaneously extract features from the 4 parts of the palmprint ROI, and then perform feature fusion and recognition, realizing high-efficiency palmprint recognition. Description of the Drawings
[0030] Figures 1a to 1b It is a schematic diagram of the palmprint ROI extraction area in the present invention, where Figure 1a is a schematic diagram of the central ROI extraction, Figure 1b is a schematic diagram of the edge ROI extraction.
[0031] Figure 2 It is a schematic diagram of the multi-input convolutional neural network structure in the present invention.
[0032] Figure 3 It is a schematic diagram of the training of the multi-input convolutional neural network of the present invention. Specific embodiments
[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0034] Refer to Figures 1a to 1b , Figure 2 , Figure 3 A high-resolution palmprint recognition method based on a multi-input convolutional neural network includes the following steps:
[0035] 1) Define the high-resolution palmprint ROI, including the cropping of the central ROI area and the edge ROI areas (including the thenar ROI area, the hypothenar ROI area, and the interdigital ROI area, abbreviated as ROI1, ROI2, ROI3). Therefore, a total of 4 ROI area images can be cropped from one palm data as the data for recognition. The steps include:
[0036] (11) The acquisition of the palmprint ROI area samples adopts an adaptive ROI cropping algorithm. First, extract the coordinates of the two valley points of the palmprint (between the middle finger and the index finger, and between the ring finger and the little finger), and then establish a two-dimensional rectangular coordinate system based on the two valley points. We use the line between the two valley points as the X-axis and establish the Y-axis perpendicular to this line, and the valley point coordinates between the ring finger and the little finger as the origin. As Figures 1a to 1b shown.
[0037] (12) For the cropping of the central ROI, according to the distance d between the two valley points, we select (-0.2d, 0.2d) as the upper left corner coordinates of the intercepted area, and then intercept 1.4d lengths along the positive directions of the X and Y coordinate axes respectively, as Figure 1a shown; for other edge ROIs, we obtain the corresponding intercepted area sizes based on the statistical analysis of the palmprint data. For ROI1, ROI2, and ROI3, we select (0.7d, 1.2d), (-0.4d, 0.7d), and (-0.25d, 0.1d) in the coordinate system as the upper left corner coordinate positions of the areas to be intercepted respectively, and then intercept 0.6d, 0.5d, and d lengths along the positive direction of the X coordinate axis respectively, and intercept 0.6d, 0.9d, and 0.4d lengths along the positive direction of the Y coordinate axis respectively. As Figure 1b shown.
[0038] (13) Therefore, the size of the central ROI area is intercepted as 1.4d * 1.4d, and the sizes of ROI1, ROI2, and ROI3 are intercepted as 0.6d * 0.6d, 0.5d * 0.9d, and d * 0.4d respectively. Then, the size of the image is normalized. Finally, the size of the central ROI is 512 * 512, and the sizes of ROI1, ROI2, and ROI3 are 512 * 512, 900 * 300, and 900 * 300 respectively. These are used as samples for training the network.
[0039] 2) Design a network model for high-resolution palmprint recognition. Use the ROI samples cropped in step 1) and perform data augmentation on them to generate more palmprint ROI data for training; the steps are as follows:
[0040] (21) Design and construct a multi-input convolutional neural network model as Figure 2 shown. Set the training parameters and loss function so that it can simultaneously extract and recognize the features of the central area and the edge area of the palmprint image. This network is mainly divided into two parts, namely the feature extraction part and the feature fusion part.
[0041] (211) The first part is the feature extraction part. The feature extraction part performs downsampling operations on the high-resolution palmprint ROI. Due to the problem of the large size difference of a set of palmprint ROI data, the feature extraction part is divided into three paths, and the central ROI, ROI2 and ROI3, and ROI1 are input respectively.
[0042] First, for the path of central ROI image feature extraction, it includes 7 layers. Each layer consists of two convolutional modules and one pooling layer. The convolutional kernel size of the convolutional layer in each convolutional module is 3 * 3, the number is 16, and the stride is 1; the stride of the pooling layer is 2; the input central ROI image passes through the first convolutional layer, BN layer, and ReLU activation function of the first convolutional module in sequence. Then, it passes through the first convolutional layer, BN layer, and ReLU activation function of the second convolutional module in sequence, and finally passes through the pooling layer for downsampling. The subsequent processing is the same as above, and it is repeated 6 times with two convolutions followed by pooling. The convolutional kernel sizes of the corresponding convolutional modules are all 3 * 3, and the numbers are 32, 64, 128, 256, 512, and 512 respectively; the central ROI image with a size of 512×512 is input and passes through the first layer to the seventh layer in sequence, and the output feature size is 4 * 4 * 512;
[0043] For the edge ROI2 and ROI3 feature extraction paths, they are similar to the previous central ROI feature extraction path. However, due to the relatively small image size, one convolutional pooling step is missing, and there are 6 layers. The number of convolutional kernels corresponding to each layer is 32, 64, 128, 256, 512, and 512 respectively. Since the sizes of the 4 ROIs vary greatly, to ensure that the extracted features can be better and more conveniently fused, during the feature extraction process, when inputting the ROI2 and ROI3 images with a size of 900*300, slicing work is performed. The image is cut into three parts with a size of 300×300, and then input sequentially; for the input of the 300×300 image, it passes through the first layer to the sixth layer in sequence, and the output feature size is 4*4*512;
[0044] The ROI1 feature extraction path is the same as the central ROI feature extraction path. The ROI1 image with a size of 512×512 is input, and it passes through the first layer to the seventh layer in sequence, and the output feature size is 4*4*512;
[0045] (212) The second part is the feature fusion part, that is, the edge ROI features and central ROI features of the palm are fused. During the above feature extraction process, the finally output feature sizes of the central ROI, ROI1, and the sliced ROI2 and ROI3 are all 4*4*512, which are consistent. Then, the features output by the three paths are cascaded for feature layer fusion. The fused features are then subjected to two convolutional modules and one pooling layer operation. The convolutional kernel size of the convolutional layer in each convolutional module is 3*3, the number is 1024, and the stride is 1; the stride of the pooling layer is 2. Finally, adaptive average pooling is performed to form a 1024-dimensional feature vector.
[0046] (22) The network training part is as Figure 3 shown. A total of three pieces of data are input, namely the anchor, positive samples, and negative samples. The multi-input convolutional neural network as Figure 2 shown is used to extract features from these three pieces of data, so their parameters are shared. Then, the chi-square distance is used to calculate the distances between the anchor and positive samples and between the anchor and negative samples, and these two distances are sent to the loss function for calculation. Then, the Adam Optimizer is used for optimization iteration, and the optimal solution is obtained after multiple iterations.
[0047] During the above training process, the loss function of the multi-input convolutional neural network is:
[0048] Loss=max(d(a,p)-d(a,n)+margin,0) (1)
[0049] Where a represents the anchor, p represents the positive sample of the same class as a, n represents the negative sample of a different class from a, margin represents a constant, set to 0.2, and d(a, n) represents the distance between the anchor a and the negative sample n. The goal of minimizing Loss is to make d(a, p) close to 0 and d(a, n) greater than d(a, p) + margin.
[0050] 3) Run the multi-input convolutional neural network model and output the results. The steps include:
[0051] (31) Run the network model as Figure 2 shown, input the test palmprint data, calculate the chi-square distance between this test data and all the reference palmprint data in the database using the trained network model, find the reference palmprint image with the minimum distance value. If the distance is less than the set threshold of 0.5, it is determined to be of the same class; otherwise, there is no class identical to the test palmprint image.
[0052] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms stated in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept of the present invention.
Claims
1. A method for high-resolution palmprint recognition based on a multi-input convolutional neural network, characterized in that, It includes the following steps: 1) Define the high-resolution palmprint ROI, including the cropping of the central ROI area and the edge ROI area. The edge ROI area includes the thenar ROI area, the hypothenar ROI area, and the interphalangeal ROI area, abbreviated as ROI1, ROI2, and ROI3; Therefore, a total of 4 ROI area images are cropped from the data of one palm as the data for recognition; 2) Design a network model for high-resolution palmprint recognition. Use the ROI samples cropped in step 1) and perform data augmentation on them to generate more palmprint ROI data for training. Specifically, it includes: (21) Design and construct a multi-input convolutional neural network model, set the training parameters and loss function, so that it can simultaneously extract and recognize the features of the central area and the edge area of the palmprint image. This multi-input convolutional neural network includes a feature extraction part and a feature fusion part; (211) The feature extraction part performs downsampling operations on the high-resolution palmprint ROI. Due to the problem of the large size difference of a set of palmprint ROI data, the feature extraction part is divided into three paths, and the central ROI, ROI2 and ROI3, and ROI1 are input respectively; First, for the path of central ROI image feature extraction, it includes 7 layers. Each layer consists of two convolutional modules and one pooling layer. The convolutional kernel size of the convolutional layer in each convolutional module is 3*3, the number is 16, and the stride is 1; the stride of the pooling layer is 2; the input central ROI image passes through the first convolutional layer, BN layer, and ReLU activation function of the first convolutional module in sequence; then passes through the first convolutional layer, BN layer, and ReLU activation function of the second convolutional module in sequence, and finally passes through the pooling layer for downsampling; then perform 6 times of two convolutional operations followed by pooling. The convolutional kernel sizes of the corresponding convolutional modules are all 3*3, and the numbers are 32, 64, 128, 256, 512, and 512 respectively; input the central ROI image with a size of 512×512, pass through the first layer to the seventh layer in sequence, and the output feature size is 4*4*512; For the edge ROI2 and ROI3 feature extraction paths, it is similar to the previous central ROI feature extraction path. However, due to the relatively small picture size, one convolutional pooling step is missing, including 6 layers. The corresponding convolutional kernel numbers in each layer are 32, 64, 128, 256, 512, and 512 respectively; since the sizes of the 4 ROIs vary greatly, in order to ensure that the extracted features can be better and more conveniently fused, slicing work is performed when inputting the ROI2 and ROI3 images with a size of 900*300; the image is cut into three parts with a size of 300×300, and then input in sequence; input the 300×300 image, pass through the first layer to the sixth layer in sequence, and the output feature size is 4*4*512; The ROI1 feature extraction path is the same as the central ROI feature extraction path. The 512×512 ROI1 image is input and passed through the first layer to the seventh layer in sequence, and the output feature size is 4*4*512; (212) The feature fusion part fuses the edge ROI feature and the central ROI feature of the palm; in the above feature extraction process, the final output feature sizes of the central ROI, ROI1, and the sliced ROI2 and ROI3 are all 4*4*512, which are kept consistent. Then, the features output by the three paths are cascaded for feature layer fusion; (22) In the network training part, three pieces of data are input, namely the anchor, the positive sample, and the negative sample. The multi-input convolutional neural network is used to extract features from these three pieces of data, so their parameters are shared; the features extracted by the network are then passed through two convolutional modules and one pooling layer operation. The convolutional kernel size of the convolutional layer in each convolutional module is 3*3, the number is 1024, and the stride is 1; the stride of the pooling layer is 2; finally, adaptive average pooling is performed to form a 1024-dimensional feature vector, and the chi-square distance is used to calculate the distances between the anchor and the positive sample and between the anchor and the negative sample, and these two distances are sent to the loss function for calculation; then, the Adam Optimizer optimizer is used for optimization iteration, and the optimal solution is obtained after multiple iterations; The loss function of the multi-input convolutional neural network is: Loss=max(d(a,p)-d(a,n)+margin,0) (1) where a represents the anchor, p represents the positive sample of the same class as a, n represents the negative sample of a different class from a, margin represents a constant, set to 0.2, and d(a,n) represents the distance between the anchor a and the negative sample n; the goal of minimizing Loss is to make d(a,p) close to 0 and d(a,n) greater than d(a,p)+margin; 3) Run the multi-input convolutional neural network model and output the results.
2. The method for high-resolution palmprint recognition based on a multi-input convolutional neural network according to claim 1, wherein The step 1) includes: (11) The acquisition of the palmprint ROI region samples adopts the adaptive ROI cropping algorithm; first, the coordinates of the two valley points of the palmprint, that is, between the middle finger and the index finger, and between the ring finger and the little finger, are extracted, and then a two-dimensional rectangular coordinate system is established based on the two valley points. The line between the two valley points is used as the X-axis, and the Y-axis is established perpendicular to this line. The valley point coordinates between the ring finger and the little finger are used as the origin; (12) For the cropping of the central ROI, according to the distance d between the two valley points, (-0.2d, 0.2d) is selected as the upper left corner coordinates of the cropping area, and then lengths of 1.4d are intercepted along the positive directions of the X and Y coordinate axes respectively; for other edge ROIs, the corresponding cropping area sizes are obtained according to the statistical analysis of the palmprint data. For ROI1, ROI2, and ROI3, (0.7d, 1.2d), (-0.4d, 0.7d), and (-0.25d, 0.1d) in the coordinate system are respectively selected as the upper left corner coordinate positions of the areas to be intercepted, and then lengths of 0.6d, 0.5d, and d are intercepted along the positive direction of the X coordinate axis respectively, and lengths of 0.6d, 0.9d, and 0.4d are intercepted along the positive direction of the Y coordinate axis respectively; (13) Therefore, the size of the central ROI area is intercepted as 1.4d * 1.4d, and the sizes of ROI1, ROI2, and ROI3 are intercepted as 0.6d * 0.6d, 0.5d * 0.9d, and d * 0.4d respectively. Then, the image size is normalized. Finally, the size of the central ROI is 512 * 512, and the sizes of ROI1, ROI2, and ROI3 are 512 * 512, 900 * 300, and 900 * 300 respectively; these are used as samples for training the network.
3. The method for high-resolution palmprint recognition based on a multi-input convolutional neural network according to claim 1 or 2, characterized in that Step 3) includes: Run the network model, input the test palmprint data, calculate the chi-square distance between this test data and all the reference palmprint data in the database using the trained network model, find the reference palmprint image with the smallest distance value. If this distance is less than the set threshold of 0.5, it is determined to be of the same class; otherwise, there is no class identical to the test palmprint image.