A palm vein recognition method, device, equipment and storage medium

Abstract

This application discloses a palm vein recognition method, apparatus, device, and storage medium. The method includes: inputting a first palm vein image to be recognized into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model; performing key point detection on the first palm vein image, and extracting first image blocks corresponding to each key point from the first palm vein image based on the key points in the first palm vein image; extracting first local feature vectors of each first image block through a pre-trained local feature extraction model; matching multiple candidate palm vein images from multiple second palm vein images through the global feature image, determining the target palm vein image from the multiple candidate palm vein images through each first local feature vector, and determining the identity information of the target palm vein image as the identity information of the first palm vein image.

Description

Technical Field

[0001] This application relates to the field of biometric technology, and in particular to a palm vein recognition method, device, equipment, and storage medium. Background Technology

[0002] With the rapid development of palm vein recognition technology, it has been widely applied to various systems for verifying personal identity, such as security systems, access control systems, subway systems, and check-in systems. Compared to other biometric technologies such as fingerprints and iris scans, palm vein recognition not only provides contactless identification but is also more difficult to forge, offering a higher level of security.

[0003] In existing technologies, global feature images are extracted from near-infrared palm vein images using global feature extraction algorithms, and then matched with palm vein images in a database to identify the user's identity information. However, global feature extraction algorithms have limited extraction capabilities and are not robust to interference conditions such as rotation and occlusion, which affects the success rate of feature matching and reduces the efficiency of palm vein recognition. Summary of the Invention

[0004] This application provides a palm vein recognition method, apparatus, device, and storage medium to quickly filter out similar candidate palm vein images through global feature images, and accurately match the target palm vein image of the same user from the candidate palm vein images through local feature vectors. This solves the problem of low feature matching success rate in the prior art and improves the efficiency of palm vein recognition.

[0005] In a first aspect, embodiments of this application provide a palm vein recognition method, including:

[0006] The first palm vein image to be identified is input into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model;

[0007] Key point detection is performed on the first palm vein image, and based on the key points in the first palm vein image, the first image block corresponding to each key point is extracted from the first palm vein image.

[0008] The first local feature vector of each of the first image patches is extracted using a pre-trained local feature extraction model.

[0009] Multiple candidate palm vein images are matched from multiple second palm vein images using the global feature image. A target palm vein image is determined from the multiple candidate palm vein images using each of the first local feature vectors. The identity information of the target palm vein image is determined to be the identity information of the first palm vein image.

[0010] Furthermore, the step of extracting first image blocks corresponding to each key point from the first palm vein image, based on the key points in the first palm vein image, includes:

[0011] The first palm vein image is divided into multiple second image blocks of a first preset size;

[0012] A first preset number of target key points are randomly selected from each key point in the first palm vein image, and each target key point is located in a different second image block;

[0013] A first image block of a first preset size is extracted from the first palm vein image, with the target key point as the center point.

[0014] Furthermore, the local feature extraction model is an encoder; correspondingly, the step of extracting the first local feature vector of each of the first image patches using the pre-trained local feature extraction model includes:

[0015] Each of the first image blocks is input into the encoder to obtain the first feature vector output by the encoder;

[0016] The pixel coordinates of the target key points are added to the first feature vector of the corresponding first image block to obtain the first local feature vector of the first image block.

[0017] Furthermore, before inputting the first palm vein image to be identified into the pre-trained global feature extraction model, the following steps are also included:

[0018] Key point detection is performed on the sample palm vein image, and based on the key points in the sample palm vein image, the first sample image block corresponding to each key point is extracted from the sample palm vein image.

[0019] The first sample image block is input into the encoder to obtain the first sample feature vector output by the encoder. The pixel coordinates of the key points are added to the first sample feature vector of the corresponding first sample image block to obtain the second sample feature vector.

[0020] The second sample feature vector is used as the third feature vector of the second sample image block where the corresponding key point is located, and the zero vector is used as the third sample feature vector of the remaining second sample image block. The second sample image block is obtained by uniformly dividing the sample palm vein image.

[0021] Each of the third sample feature vectors is input into the decoder to obtain the decoded image output by the decoder;

[0022] Based on the reconstruction loss function, a first loss value is determined using the decoded image and the sample palm vein image, and the network parameters of the decoder and the encoder are optimized using the first loss value.

[0023] Furthermore, when determining a first loss value based on the reconstruction loss function using the decoded image and the sample palm vein image, and optimizing the network parameters of the decoder and the encoder using the first loss value, the method further includes:

[0024] Based on the nearest neighbor metric loss function, a second loss value is determined using the second sample feature vector, and the network parameters of the encoder are optimized using the second loss value.

[0025] Furthermore, the step of determining the second loss value based on the nearest neighbor metric loss function using the second sample feature vector includes:

[0026] Obtain a third palm vein image belonging to the same user as the sample palm vein image, perform key point detection on the third palm vein image, and extract the third image block corresponding to each key point from the third palm vein image based on the key points in the third palm vein image;

[0027] The third image block is input into the encoder to obtain the second feature vector output by the encoder. The pixel coordinates of the key points are added to the second feature vector of the corresponding third image block to obtain the third feature vector.

[0028] Based on the feature distance between the second sample feature vector and each of the third feature vectors, the third feature vector that is closest to the second sample feature vector is taken as a feature vector of the same class, and the third feature vector that is second closest to the second sample feature vector is taken as a feature vector of a different class.

[0029] Based on the nearest neighbor metric loss function, the second loss value is determined using each of the second feature vectors and the corresponding similar and dissimilar feature vectors.

[0030] Furthermore, determining the target palm vein image from the plurality of candidate palm vein images using each of the first local feature vectors includes:

[0031] Each of the first local feature vectors is matched with each of the second local feature vectors of the candidate palm vein image;

[0032] If the matching degree between the first local feature vector and the second local feature vector is greater than or equal to a preset matching threshold, it is determined that the key point corresponding to the first local feature vector matches the key point of the second local feature vector.

[0033] The candidate palm vein image with the highest number of key point matches is selected as the target palm vein image.

[0034] Secondly, embodiments of this application provide a palm vein recognition device, comprising:

[0035] The global feature extraction module is configured to input the first palm vein image to be identified into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model.

[0036] The image block cropping module is configured to perform key point detection on the first palm vein image and, based on the key points in the first palm vein image, crop the first image block corresponding to each key point from the first palm vein image.

[0037] The local feature extraction module is configured to extract the first local feature vector of each of the first image patches using a pre-trained local feature extraction model.

[0038] The identity recognition module is configured to match multiple candidate palm vein images from multiple second palm vein images using the global feature image, determine a target palm vein image from the multiple candidate palm vein images using each of the first local feature vectors, and determine the identity information of the target palm vein image as the identity information of the first palm vein image.

[0039] Furthermore, the image patch extraction module includes:

[0040] The first image segmentation unit is configured to divide the first palm vein image into multiple second image blocks of a first preset size;

[0041] The first key point selection unit is configured to randomly select a first preset number of target key points from each key point of the first palm vein image, wherein each target key point is located in a different second image block.

[0042] The first image block cropping unit is configured to crop a first image block of a first preset size from the first palm vein image, with the target key point as the center point.

[0043] Furthermore, the local feature extraction model is an encoder; correspondingly, the local feature extraction module includes:

[0044] The first vector extraction unit is configured to input each first image block into the encoder to obtain a first feature vector output by the encoder.

[0045] The first local feature generation unit is configured to add the pixel coordinates of the target key point to the first feature vector of the corresponding first image block to obtain the first local feature vector of the first image block.

[0046] Furthermore, the palm vein recognition device also includes a model training module, which includes:

[0047] The second image block cropping unit is configured to perform key point detection on the sample palm vein image and, based on the key points in the sample palm vein image, crop out the first sample image block corresponding to each key point from the sample palm vein image.

[0048] The second vector extraction unit is configured to input the first sample image block into the encoder to obtain the first sample feature vector output by the encoder, and add the pixel coordinates of the key points into the first sample feature vector of the corresponding first sample image block to obtain the second sample feature vector.

[0049] The third vector generation unit is configured to use the second sample feature vector as the third feature vector of the second sample image block where the corresponding key point is located, and use the zero vector as the third sample feature vector of the remaining second sample image block, wherein the second sample image block is obtained by uniformly dividing the sample palm vein image.

[0050] The decoding unit is configured to input the feature vectors of each of the third samples into the decoder to obtain the decoded image output by the decoder;

[0051] The first optimization unit is configured to determine a first loss value based on the decoded image and the sample palm vein image using a reconstruction loss function, and to optimize the network parameters of the decoder and the encoder using the first loss value.

[0052] Furthermore, the model training module includes:

[0053] The second optimization unit is configured to, when determining a first loss value based on the reconstruction loss function using the decoded image and the sample palm vein image, and optimizing the network parameters of the decoder and the encoder using the first loss value, determine a second loss value based on the nearest neighbor metric loss function using the second sample feature vector, and optimize the network parameters of the encoder using the second loss value.

[0054] Furthermore, the second optimization unit includes:

[0055] The third image block cropping subunit is configured to acquire a third palm vein image belonging to the same user as the sample palm vein image, perform key point detection on the third palm vein image, and crop the third image block corresponding to each key point from the third palm vein image based on the key points in the third palm vein image.

[0056] The third vector extraction subunit is configured to input the third image block into the encoder to obtain the second feature vector output by the encoder, and add the pixel coordinates of the key points to the second feature vector of the corresponding third image block to obtain the third feature vector;

[0057] The triplet construction subunit is configured to, based on the feature distance between the second sample feature vector and each of the third feature vectors, take the third feature vector that is closest to the second sample feature vector as the same type of feature vector, and take the third feature vector that is second closest to the second sample feature vector as the different type of feature vector.

[0058] The second loss value determination subunit is configured to determine the second loss value based on the nearest neighbor metric loss function, using each of the second feature vectors and the corresponding similar and dissimilar feature vectors.

[0059] Furthermore, the identity recognition module includes:

[0060] The feature vector matching unit is configured to match each of the first local feature vectors with each of the second local feature vectors of the candidate palm vein image;

[0061] The key point matching unit is configured to determine that the key point corresponding to the first local feature vector matches the key point of the second local feature vector when the matching degree between the first local feature vector and the second local feature vector is greater than or equal to a preset matching threshold.

[0062] The target determination unit is configured to select the candidate palm vein image with the most key point matches as the target palm vein image.

[0063] Thirdly, embodiments of this application provide a palm vein recognition device, including:

[0064] Memory and one or more processors;

[0065] The memory is used to store one or more programs;

[0066] When the one or more programs are executed by the one or more processors, the one or more processors implement the palm vein recognition method as described in the first aspect.

[0067] Fourthly, embodiments of this application provide a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the palm vein recognition method as described in the first aspect.

[0068] In this application, a first palm vein image to be identified is input into a pre-trained global feature extraction model to obtain a global feature image output by the model. Keypoint detection is performed on the first palm vein image, and first image blocks corresponding to each keypoint are extracted from the first palm vein image. A first local feature vector is extracted from each first image block using a pre-trained local feature extraction model. Multiple candidate palm vein images are matched from multiple second palm vein images using the global feature image, and the target palm vein image is determined from the multiple candidate palm vein images using each first local feature vector. The identity information of the target palm vein image is then used as the identity information of the first palm vein image. Through these techniques, similar candidate palm vein images are quickly filtered out using the global feature image, and the target palm vein image of the same user is accurately matched from the candidate palm vein images using local feature vectors. This improves the success rate of feature matching, solves the problem of low feature matching success rate in existing technologies, and improves the efficiency of palm vein recognition. Attached Figure Description

[0069] Figure 1 This is a flowchart of a palm vein recognition method provided in one embodiment of this application;

[0070] Figure 2 This is a flowchart of the process of capturing a first image block provided in an embodiment of this application;

[0071] Figure 3 This is a flowchart of the extraction of the first local feature vector provided in an embodiment of this application;

[0072] Figure 4 This is a flowchart of the training of the local feature extraction model provided in the embodiments of this application;

[0073] Figure 5 This is a flowchart of determining the second loss value provided in an embodiment of this application;

[0074] Figure 6 This is a flowchart illustrating the determination of a target palm vein image provided in an embodiment of this application;

[0075] Figure 7 This is a schematic diagram of the structure of a palm vein recognition device provided in one embodiment of this application;

[0076] Figure 8 This is a schematic diagram of the structure of a palm vein recognition device provided in one embodiment of this application. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0078] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0079] The palm vein recognition method, apparatus, device, and storage medium provided in this application aim to input a first palm vein image to be recognized into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model; perform key point detection on the first palm vein image, and extract first image blocks corresponding to each key point from the first palm vein image based on the key points in the first palm vein image; extract first local feature vectors for each first image block through a pre-trained local feature extraction model; match multiple candidate palm vein images from multiple second palm vein images using the global feature image; determine the target palm vein image from the multiple candidate palm vein images using each of the first local feature vectors; and determine the identity information of the target palm vein image as the identity information of the first palm vein image. Compared with traditional palm vein recognition methods, its global feature extraction algorithm extracts a global feature image from near-infrared palm vein images and matches the global feature image with palm vein images in a database to identify the user's identity information. However, the extraction capability of the global feature extraction algorithm is limited, and its robustness to interference conditions such as rotation and occlusion is weak, affecting the success rate of feature matching and resulting in reduced efficiency of palm vein recognition. Based on this, the palm vein recognition method, device, equipment, and storage medium proposed in the embodiments of this application can solve the problem of low feature matching success rate in the prior art.

[0080] Figure 1 This is a flowchart of a palm vein recognition method provided in one embodiment of this application. The palm vein recognition method provided in this embodiment can be executed by a palm vein recognition device, which can be implemented by software and / or hardware. The palm vein recognition device can be composed of two or more physical entities, or it can be composed of a single physical entity.

[0081] The following description uses a palm vein recognition device as the main entity performing the palm vein recognition method. (Reference) Figure 1 The palm vein recognition method includes:

[0082] S110. Input the first palm vein image to be identified into the pre-trained global feature extraction model to obtain the global feature image output by the global feature extraction model.

[0083] This embodiment describes a check-in system equipped with a palm vein recognition device as an example. In this embodiment, the palm vein recognition device includes a near-infrared camera, which captures a near-infrared palm vein image of the user. The first palm vein image can be understood as the near-infrared palm vein image of the user to be checked in captured by the near-infrared camera. The palm vein recognition device can determine the user's identity information based on the first palm vein image, thereby completing the user's check-in.

[0084] In this embodiment, the global feature extraction model is a pre-trained neural network model used to extract overall feature information from the first palm vein image, obtaining a global feature image of the first palm vein image. In one embodiment, considering that near-infrared palm vein images have significant local structure and detail information, a VIT model can be used as the global feature extraction model. The VIT model adopts a ViT-Base structure, and the network parameters of the VIT model are trained based on an identity loss function. For example, through the VIT model, the first palm vein image can be divided into multiple slice images of a second preset size, and each slice image is converted into a one-dimensional vector. Each one-dimensional vector is input into a linear projection layer to obtain the first slice feature vectors output by the linear projection layer. The first slice feature vectors are summed with the preset position encoding to obtain the second slice feature vectors. The second slice feature vectors are input into a transformer encoder. The transformer encoder extracts image feature sequences from the second slice feature vectors based on a multi-head self-attention model. The image feature sequences are input into a fully connected layer to obtain the global feature image output by the fully connected layer. Compared to traditional convolutional models, the multi-head self-attention model in the VIT model can not only capture the relationship features between different slice images, but also detect some local features individually, and then accumulate them in the final global feature image, which helps to improve the accuracy of the global feature image. Moreover, as a lightweight model, the VIT model requires less runtime resources and has high feature extraction efficiency, which is beneficial for optimizing the efficiency of palm vein recognition.

[0085] S120. Perform key point detection on the first palm vein image, and extract the first image block corresponding to each key point from the first palm vein image based on the key points in the first palm vein image.

[0086] Although global feature extraction is fast, its robustness to external interference factors such as image rotation and occlusion is poor, easily leading to errors in user identification and a low success rate of feature matching. Therefore, this embodiment proposes to extract local features of key points in the first palm vein image to reduce the influence of factors such as rotation and occlusion, thereby obtaining robust identification results and improving the success rate of feature matching.

[0087] In this embodiment, before extracting the local features of key points, the first palm vein image is first detected using a SIFT (Scale-invariant feature transform) detector to obtain key points in the first palm vein image. Using the key points as the center points of the corresponding first image blocks, a first image block of a first preset size is extracted from the first palm vein image, and the local features of the corresponding key points are extracted from the first image blocks corresponding to the key points.

[0088] Because the number of key points in the first palm vein image is large, feature extraction from a large number of first image blocks is inefficient, affecting the efficiency of palm vein recognition. If the distribution distance of some key points is small, the first image blocks of some key points will largely overlap. The local features extracted from the largely overlapping first image blocks will have high similarity, meaning that some extracted local features are invalid. To improve the efficiency of palm vein recognition, the number of generated first image blocks can be reduced, and spatial constraints can be added to the key points corresponding to each first image block. This will allow the key points of each first image block to be distributed more widely, so that the generated first image blocks can cover a wider range with lower overlap, ensuring the effectiveness of the extracted local features. In this embodiment, Figure 2 This is a flowchart illustrating the process of cropping a first image block, as provided in an embodiment of this application. Figure 2 As shown, the step of cropping the first image block specifically includes S1201-S1203:

[0089] S1201, Divide the first palm vein image into multiple second image blocks of a first preset size.

[0090] S1202. Randomly select a first preset number of target key points from each key point of the first palm vein image, wherein each target key point is located in a different second image block.

[0091] S1203. Extract a first image block of a first preset size from the first palm vein image, with the target key point as the center point.

[0092] For example, suppose the size of the first palm vein image is 224x224, the first preset size is 32x32, and the first preset number is 20. The first palm vein image is divided into 7x7 non-overlapping regions. Each region is treated as a 32x32 second image block. Twenty target keypoints are randomly selected from all detected keypoints, each falling into a different second image block, to ensure wider keypoint coverage. Then, using the target keypoints as the center points of the corresponding first image blocks, 32x32 first image blocks are cropped from the first palm vein image, resulting in 20 first image blocks centered on the corresponding target keypoints.

[0093] S130. Extract the first local feature vector of each first image patch through a pre-trained local feature extraction model.

[0094] In this embodiment, the local feature extraction model is a pre-trained neural network model used to extract local feature information of the first image patch for each target key point, thereby obtaining the first local feature vector of each first image patch. In one embodiment, an encoder is used as the local feature extraction model to extract the local features of each first image patch. Figure 3This is a flowchart illustrating the extraction of the first local feature vector provided in an embodiment of this application. For example... Figure 3 As shown, the step of extracting the first local feature vector specifically includes S1301-S1302:

[0095] S1301. Input each first image block into the encoder to obtain the first feature vector output by the encoder.

[0096] S1302. Add the pixel coordinates of the target key points to the first feature vector of the corresponding first image block to obtain the first local feature vector of the first image block.

[0097] For example, the encoder includes a three-layer convolutional network and a fully connected layer. A 32x32 first image patch is input into the first-layer 3x3 convolutional network to obtain a 16x16x32 first feature image. This 16x16x32 first feature image is then input into the second-layer 3x3 convolutional network to obtain an 8x8x64 second feature image. This 8x8x64 second feature image is then input into the third-layer 3x3 convolutional network to obtain a 4x4x128 third feature image. This third feature image is then input into the fully connected layer to obtain a 126-dimensional first feature vector. The pixel coordinates of the target keypoints corresponding to the first image patch in the first palm vein image are added to this first feature vector to obtain a 128-dimensional first local feature vector. This process is repeated, with each first image patch sequentially input into the encoder, and the pixel coordinates of the corresponding target keypoints added to the encoder's output first feature vector to obtain the first local feature vector corresponding to each first image patch.

[0098] In this embodiment, Figure 4 This is a flowchart of the training of the local feature extraction model provided in the embodiments of this application. For example... Figure 4 As shown, the specific steps for training the local feature extraction model include S1501-S1505:

[0099] S1501. Perform key point detection on the sample palm vein image. Based on the key points in the sample palm vein image, extract the first sample image block corresponding to each key point from the sample palm vein image.

[0100] The sample palm vein images are used as sample data for training the local feature extraction model. For example, a sample dataset is pre-collected, comprising multiple sample data points. A 224x224 sample palm vein image is obtained from the pre-collected dataset. Keypoints are extracted from the sample palm vein images using the SIFI detector. The 224x224 sample palm vein image is divided into 49 non-overlapping 32x32 second sample image blocks. From all detected keypoints, 20 keypoints falling into different second sample image blocks are randomly selected. Using these keypoints as the center point of the corresponding first sample image block, a 32x32 first sample image block is extracted from the sample palm vein image, resulting in 20 first sample image blocks.

[0101] S1502. Input the first sample image block into the encoder to obtain the first sample feature vector output by the encoder. Add the pixel coordinates of the key points to the first sample feature vector of the corresponding first sample image block to obtain the second sample feature vector.

[0102] For example, a 32x32 first sample image block is input into the encoder to obtain a 126-dimensional first sample feature vector output by the encoder. The pixel coordinates of the key points of the first sample image block in the sample palm vein image are added to the first sample feature vector to obtain a 128-dimensional second sample feature vector. By analogy, 20 128-dimensional second sample feature vectors can be obtained.

[0103] S1503. The second sample feature vector is used as the third feature vector of the second sample image block where the corresponding key point is located, and the zero vector is used as the third sample feature vector of the remaining second sample image blocks. The second sample image blocks are obtained by uniformly dividing the sample palm vein image.

[0104] For example, the input to the decoder is 49x128 feature information, that is, the local features of 49 regions of the sample palm vein image. The second sample feature vector can be used as the local feature of the region where the corresponding key point belongs, that is, the third sample feature vector. In other words, 20 128-dimensional second sample feature vectors can occupy 20 regions, and the local features of the remaining 29 regions, that is, the third sample feature vectors, adopt 128-dimensional zero vectors.

[0105] S1504. Input the feature vectors of each third sample into the decoder to obtain the decoded image output by the decoder.

[0106] For example, the third sample feature vector of each region is used as the input to the corresponding region location of the decoder. The decoder uses a three-layer 3x3 deconvolution layer. After the third sample feature vector of each region is input into the decoder, the decoder outputs a 224x224 decoded image.

[0107] S1505. Based on the reconstruction loss function, a first loss value is determined by decoding the image and the sample palm vein image, and the network parameters of the decoder and encoder are optimized by the first loss value.

[0108] In this embodiment, the purpose of the reconstruction loss is to ensure that the decoded image generated by the decoder is as consistent as possible with the sample palm vein image. For the decoder, the input is 49x128 feature information, and the output is a 224x224 decoded image, where 49 represents the region blocks at each location, and 128 represents the dimension of the local features of the region blocks. For the local features of the region blocks that detect 20 key points, the local features of the key points are replaced by the second sample feature vector of the first sample image block corresponding to the key points, while the local features of the remaining region blocks use zero vectors. This input construction results in a large amount of blank content in the decoder's input, simulating the situation where key information is damaged or occluded in near-infrared palm vein images. Furthermore, the relatively sparse input allows the encoder to learn more contextual and palm vein structure information during feature learning. Applying reconstruction constraints under conditions of largely missing information improves the robustness of the network output, ensuring that the local feature vectors extracted by the encoder have rich contextual semantic information, effectively improving the accuracy of the local feature vectors extracted by the local feature extraction model.

[0109] In this embodiment, assuming the sample palm vein image is x and the decoded image is z, the expression for the reconstruction loss function is as follows:

[0110]

[0111] For example, the first loss value can be calculated using the above expression for the reconstruction loss function, and the network parameters of the encoder and decoder are optimized based on the first loss value. The encoder and decoder are trained based on multiple sample palm vein images. When the encoder and decoder models converge or the training iterations reach the upper limit, the training of the encoder and decoder is completed, and the trained encoder is used as a local feature extraction model.

[0112] Based on optimizing the network parameters of the decoder and encoder using the reconstruction loss function, a second loss value can be determined using the second sample feature vector based on the nearest neighbor metric loss function. This second loss value can then be used to optimize the encoder's network parameters. For example, the nearest neighbor metric loss function can calculate the local feature distance between matching or non-matching keypoints in two near-infrared palm vein images belonging to the same user. Optimizing the encoder's network parameters based on this local feature distance allows the encoder to learn discriminative information, thereby better describing the structure and texture details of the first image patch of the target keypoint and improving the accuracy of the local feature vector.

[0113] While metric learning is an effective learning strategy to improve the discriminative ability of a network, it requires labeling the learning objects. During feature learning, the encoder needs to label each pair of keypoints between different sample palm vein images to determine if they match, which is extremely time-consuming and labor-intensive because the number of keypoints is tens of times greater than the number of images. To reduce the workload of labeling corresponding keypoints, this embodiment proposes an unsupervised learning strategy to improve model training efficiency. In this embodiment, Figure 5 This is a flowchart for determining the second loss value provided in an embodiment of this application. For example... Figure 5 As shown, the step of determining the second loss value specifically includes S1506-S1509:

[0114] S1506. Obtain a third palm vein image belonging to the same user as the sample palm vein image, perform key point detection on the third palm vein image, and extract the third image block corresponding to each key point from the third palm vein image based on the key points in the third palm vein image.

[0115] The third palm vein image can be a sample palm vein image from the dataset, or a near-infrared palm vein image collected from the user corresponding to the sample palm vein image. Keypoints are extracted from the third palm vein image using the SIFI detector. The 224x224 third palm vein image is divided into 49 non-overlapping 32x32 fourth image blocks. Twenty keypoints falling into different fourth image blocks are randomly selected from all keypoints in the third palm vein image, and these keypoints are used as the center points of the third image blocks. A 32x32 third image block is then extracted from the third palm vein image.

[0116] S1507. Input the third image block into the encoder to obtain the second feature vector output by the encoder. Add the pixel coordinates of the key points to the second feature vector of the corresponding third image block to obtain the third feature vector.

[0117] For example, 20 third image blocks are sequentially input into the current encoder to obtain the second feature vector corresponding to each third image block output by the encoder. The pixel coordinates of the key points of each third image block in the third palm vein image are added to the corresponding second feature vector to obtain the third feature vector.

[0118] S1508. Based on the feature distance between the second sample feature vector and each third feature vector, the third feature vector that is closest to the second sample feature vector is taken as the same type of feature vector, and the third feature vector that is second closest to the second sample feature vector is taken as the different type of feature vector.

[0119] For example, the second sample feature vector can represent the local features of key points in the sample palm vein image, and the third feature vector can represent the local features of key points in the third palm vein image. The local features of key points contain 126-dimensional visual information and 2-dimensional spatial information. If a set of second sample feature vectors and third feature vectors are closest to each other in terms of feature distance with all second sample feature vectors or all third feature vectors, it indicates that the key points corresponding to the second sample feature vectors and the third feature vectors are matched, and the third feature vectors of this set can be used as the same type of feature vector as the corresponding second sample feature vector. At the same time, the key points of the second sample feature vectors can be considered as outliers compared to the key points of other third feature vectors. To avoid overly complex training, a hard-example strategy is adopted, using the third feature vector corresponding to the outlier point with the closest feature distance to the second sample feature vector as the outlier feature vector. Triples are constructed by combining the second sample feature vectors with their corresponding same-type and outlier feature vectors to achieve the goal of minimizing the same-type distance and maximizing the outlier distance through these triples.

[0120] For example, the second sample feature vector in the sample palm vein image includes Where f i a This refers to the second sample feature vector of the i-th keypoint in the sample palm vein image. The third feature vector in the third palm vein image includes... Where f i b This refers to the third feature vector of the i-th key point in the sample palm vein image. If and Of all feature distances, the closest feature distance is and Of all feature distances, the closest feature distance is Then it can be determined and They are closest to each other, and thus will As Similar eigenvectors. If and The feature distance is second only to and The feature distance, then As The heterogeneous feature vectors are then used to construct triples. By analogy, all triples in the sample palm vein image and the third palm vein image are identified.

[0121] S1509. Based on the nearest neighbor metric loss function, determine the second loss value through each second feature vector and its corresponding similar and dissimilar feature vectors.

[0122] For example, the nearest neighbor metric loss function uses the triplet loss function, the expression of which is shown below:

[0123]

[0124] Among them, f a f is the feature vector of the second sample in the triplet. p f represents the eigenvectors of the same type in the triplet. n Let be the out-of-class feature vector in the triplet. Substituting the triplet into the expression of the triplet loss function yields the second loss value.

[0125] S140. Multiple candidate palm vein images are matched from multiple second palm vein images using global feature images. The target palm vein image is determined from multiple candidate palm vein images using each first local feature vector. The identity information of the target palm vein image is determined as the identity information of the first palm vein image.

[0126] For example, the palm vein recognition device pre-stores second palm vein images of all users, with each second palm vein image associated with the corresponding user's identity information. The global feature image is sequentially matched with each pre-stored second palm vein image to find the top 10 most similar second palm vein images as candidate palm vein images. It is understood that the number of first local feature vectors in the first palm vein image and the number of second local feature vectors in the second palm vein image are large. If the first local feature vector of the first palm vein image is matched one by one with the second local feature vector of each second palm vein image, the matching efficiency is low, affecting the palm vein recognition efficiency. Therefore, multiple similar candidate palm vein images can be quickly filtered from all second palm vein images using the global feature image, and then the target palm vein image belonging to the same user as the first palm vein image can be further determined using the first local feature vector. This ensures palm vein recognition accuracy while effectively improving palm vein recognition efficiency. The second local feature vector is a local feature extracted from the fifth image patch of each key point of the second palm vein image using a pre-trained local feature extraction model. The process of determining the fifth image block is the same as that of determining the first image block; please refer to step S120 for details.

[0127] In this embodiment, Figure 6 This is a flowchart illustrating the determination of a target palm vein image provided in an embodiment of this application. For example... Figure 6 As shown, the steps for determining the target palm vein image specifically include S1401-S1403:

[0128] S1401. Match each first local feature vector with each second local feature vector of the candidate palm vein image.

[0129] S1402. If the matching degree between the first local feature vector and the second local feature vector is greater than or equal to a preset matching threshold, determine that the key point corresponding to the first local feature vector matches the key point of the second local feature vector.

[0130] S1403. Select the candidate palm vein image with the most key point matching as the target palm vein image.

[0131] For example, the preset matching threshold is the minimum matching degree of the corresponding local features when two key points are mutually matched points, as set in this embodiment. The first local feature vector can be regarded as the local feature of the target key point corresponding to the first image block, and the second local feature vector can be regarded as the local feature of the key point corresponding to the fifth image block. The matching degree of the first local feature vector with each of the second local feature vectors of a candidate palm vein image is calculated respectively. When the matching degree of the first local feature vector with a certain second local feature vector is greater than or equal to the preset matching threshold, it indicates that the key point of the first local feature vector and the key point of the second local feature vector are mutually matched points. When the matching degree of the first local feature vector with a certain second local feature vector is less than the preset matching threshold, it indicates that the key point of the first local feature vector and the key point of the second local feature vector are not matched points. When the matching degree of the first local feature vector with multiple second local feature vectors is greater than or equal to the preset matching threshold, the key point of the second local feature vector with the largest matching degree is selected as the matching point of the key point of the first local feature vector. In this way, the number of matched key points in the first palm vein image and the candidate palm vein image is determined. The more key points matched, the higher the similarity between the first palm vein image and the corresponding candidate palm vein image. Therefore, the candidate palm vein image with the most key point matches can be used as the target palm vein image, and the identity information associated with and saved with the target palm vein image can be used as the identity information of the user of the first palm vein image.

[0132] In summary, the palm vein recognition method provided in this application involves inputting a first palm vein image to be recognized into a pre-trained global feature extraction model to obtain a global feature image output by the model; performing key point detection on the first palm vein image and extracting first image blocks corresponding to each key point based on the key points in the first palm vein image; extracting first local feature vectors for each first image block using a pre-trained local feature extraction model; matching multiple candidate palm vein images from multiple second palm vein images using the global feature image; determining the target palm vein image from the multiple candidate palm vein images using each first local feature vector; and identifying the identity information of the target palm vein image as the identity information of the first palm vein image. Through these technical means, similar candidate palm vein images are quickly filtered out using the global feature image, and the target palm vein image of the same user is accurately matched from the candidate palm vein images using local feature vectors, improving the success rate of feature matching, solving the problem of low feature matching success rate in existing technologies, and improving the efficiency of palm vein recognition.

[0133] Based on the above embodiments, Figure 7 This is a schematic diagram of the structure of a palm vein recognition device according to one embodiment of this application. (Reference) Figure 7 The palm vein recognition device provided in this embodiment specifically includes: a global feature extraction module 21, an image block cropping module 22, a local feature extraction module 23, and an identity recognition module 24.

[0134] The global feature extraction module 21 is configured to input the first palm vein image to be identified into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model.

[0135] The image block cropping module 22 is configured to perform key point detection on the first palm vein image and crop the first image block corresponding to each key point from the first palm vein image based on the key points in the first palm vein image.

[0136] The local feature extraction module 23 is configured to extract the first local feature vector of each first image patch through a pre-trained local feature extraction model;

[0137] The identity recognition module 24 is configured to match multiple candidate palm vein images from multiple second palm vein images using global feature images, determine the target palm vein image from multiple candidate palm vein images using each first local feature vector, and determine the identity information of the target palm vein image as the identity information of the first palm vein image.

[0138] Based on the above embodiments, the image block cropping module 22 includes:

[0139] The first image segmentation unit is configured to divide the first palm vein image into multiple second image blocks of a first preset size;

[0140] The first key point selection unit is configured to randomly select a first preset number of target key points from each key point of the first palm vein image, wherein each target key point is located in a different second image block.

[0141] The first image block cropping unit is configured to crop a first image block of a first preset size from the first palm vein image, with the target key point as the center point.

[0142] Based on the above embodiments, the local feature extraction model is an encoder; correspondingly, the local feature extraction module 23 includes:

[0143] The first vector extraction unit is configured to input each first image block into the encoder to obtain the first feature vector output by the encoder.

[0144] The first local feature generation unit is configured to add the pixel coordinates of the target key points to the first feature vector of the corresponding first image block to obtain the first local feature vector of the first image block.

[0145] Based on the above embodiments, the palm vein recognition device further includes a model training module, which includes:

[0146] The second image block cropping unit is configured to perform key point detection on the sample palm vein image and, based on the key points in the sample palm vein image, crop the first sample image block corresponding to each key point from the sample palm vein image.

[0147] The second vector extraction unit is configured to input the first sample image block into the encoder to obtain the first sample feature vector output by the encoder, and add the pixel coordinates of the key points into the first sample feature vector of the corresponding first sample image block to obtain the second sample feature vector.

[0148] The third vector generation unit is configured to use the second sample feature vector as the third feature vector of the second sample image block where the corresponding key point is located, and use the zero vector as the third sample feature vector of the remaining second sample image block. The second sample image block is obtained by uniformly dividing the sample palm vein image.

[0149] The decoding unit is configured to input the feature vectors of each third sample into the decoder to obtain the decoded image output by the decoder;

[0150] The first optimization unit is configured to determine a first loss value based on the reconstruction loss function by decoding the image and the sample palm vein image, and to optimize the network parameters of the decoder and encoder based on the first loss value.

[0151] Based on the above embodiments, the model training module includes:

[0152] The second optimization unit is configured to determine a second loss value based on the nearest neighbor metric loss function using the second sample feature vector, and to optimize the network parameters of the encoder using the second loss value when determining a first loss value based on the reconstruction loss function using the decoded image and the sample palm vein image, and optimizing the network parameters of the encoder using the second loss value.

[0153] Based on the above embodiments, the second optimization unit includes:

[0154] The third image block cropping subunit is configured to acquire a third palm vein image belonging to the same user as the sample palm vein image, perform key point detection on the third palm vein image, and crop the third image block corresponding to each key point from the third palm vein image based on the key points in the third palm vein image.

[0155] The third vector extraction subunit is configured to input the third image block into the encoder to obtain the second feature vector output by the encoder, and add the pixel coordinates of the key points to the second feature vector of the corresponding third image block to obtain the third feature vector.

[0156] The triplet construction subunit is configured to, based on the feature distance between the second sample feature vector and each third feature vector, take the third feature vector that is closest to the second sample feature vector as the same type of feature vector, and take the third feature vector that is second closest to the second sample feature vector as the different type of feature vector.

[0157] The second loss value determination subunit is configured to determine the second loss value based on the nearest neighbor metric loss function, using each second feature vector and its corresponding similar and dissimilar feature vectors.

[0158] Based on the above embodiments, the identity recognition module 24 includes:

[0159] The feature vector matching unit is configured to match each first local feature vector with each second local feature vector of the candidate palm vein image;

[0160] The key point matching unit is configured to determine that the key points corresponding to the first local feature vector match the key points of the second local feature vector when the matching degree between the first local feature vector and the second local feature vector is greater than or equal to a preset matching threshold.

[0161] The target determination unit is configured to select the candidate palm vein image with the highest number of keypoint matches as the target palm vein image.

[0162] In summary, the palm vein recognition device provided in this application provides the following steps: First, a first palm vein image to be recognized is input into a pre-trained global feature extraction model to obtain a global feature image output by the model. Key point detection is performed on the first palm vein image, and first image blocks corresponding to each key point are extracted from the first palm vein image. First local feature vectors for each first image block are extracted using a pre-trained local feature extraction model. Multiple candidate palm vein images are matched from multiple second palm vein images using the global feature image. The target palm vein image is determined from the multiple candidate palm vein images using each first local feature vector, and the identity information of the target palm vein image is identified as the identity information of the first palm vein image. Through these technical means, similar candidate palm vein images are quickly filtered out using the global feature image, and the target palm vein image of the same user is accurately matched from the candidate palm vein images using local feature vectors, improving the success rate of feature matching and solving the problem of low feature matching success rate in the prior art, thus improving the efficiency of palm vein recognition.

[0163] The palm vein recognition device provided in this application embodiment can be used to execute the palm vein recognition method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0164] This application provides a palm vein recognition device, referring to... Figure 8 The palm vein recognition device includes a processor 31, a memory 32, a communication device 33, an input device 34, and an output device 35. The number of processors 31 and the number of memories 32 in the palm vein recognition device can be one or more. The processor 31, memory 32, communication device 33, input device 34, and output device 35 of the palm vein recognition device can be connected via a bus or other means.

[0165] The memory 32, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the palm vein recognition method described in any embodiment of this application (e.g., the global feature extraction module 21, image block capture module 22, local feature extraction module 23, and identity recognition module 24 in the palm vein recognition device). The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created based on the use of the device, etc. Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, which can be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0166] The communication device 33 is used for data transmission.

[0167] The processor 31 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory, thereby realizing the palm vein recognition method described above.

[0168] Input device 34 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 35 may include display devices such as a display screen.

[0169] The palm vein recognition device provided above can be used to execute the palm vein recognition method provided in the above embodiments, and has corresponding functions and beneficial effects.

[0170] This application embodiment also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to execute a palm vein recognition method. The palm vein recognition method includes: inputting a first palm vein image to be recognized into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model; performing keypoint detection on the first palm vein image, and extracting first image blocks corresponding to each keypoint from the first palm vein image based on the keypoints in the first palm vein image; extracting first local feature vectors for each first image block using a pre-trained local feature extraction model; matching multiple candidate palm vein images from multiple second palm vein images using the global feature image; determining a target palm vein image from the multiple candidate palm vein images using each first local feature vector; and determining the identity information of the target palm vein image as the identity information of the first palm vein image.

[0171] Storage medium – any type of memory device or storage device. The term “storage medium” is intended to include: mounting media, such as CD-ROM, floppy disk, or magnetic tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements, etc. Storage medium may also include other types of memory or combinations thereof. Furthermore, storage medium may reside in a first computer system in which the program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term “storage medium” can include two or more storage media residing in different locations (e.g., in different computer systems connected via a network). Storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0172] Of course, the computer-executable instructions provided in the embodiments of this application are not limited to the palm vein recognition method described above, but can also perform related operations in the palm vein recognition method provided in any embodiment of this application.

[0173] The palm vein recognition device, equipment, and storage medium provided in the above embodiments can execute the palm vein recognition method provided in any embodiment of this application. For technical details not described in detail in the above embodiments, please refer to the palm vein recognition method provided in any embodiment of this application.

[0174] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the claims.

Claims

1. A method for palm vein recognition, characterized in that, include: The first palm vein image to be identified is input into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model; Key point detection is performed on the first palm vein image. Based on the key points in the first palm vein image, a first image block corresponding to each key point is extracted from the first palm vein image. The step of extracting the first image block corresponding to each key point from the first palm vein image includes: dividing the first palm vein image into multiple second image blocks of a first preset size; randomly selecting a first preset number of target key points from each key point in the first palm vein image, with each target key point located in a different second image block; and extracting a first image block of a first preset size centered on the target key point from the first palm vein image. The first local feature vector of each of the first image patches is extracted using a pre-trained local feature extraction model. Multiple candidate palm vein images are matched from multiple second palm vein images using the global feature image, and a target palm vein image is determined from the multiple candidate palm vein images using each of the first local feature vectors. The identity information of the target palm vein image is determined to be the identity information of the first palm vein image. Before inputting the first palm vein image to be identified into the pre-trained global feature extraction model, the method further includes: Key point detection is performed on the sample palm vein image. Based on the key points in the sample palm vein image, the first sample image block corresponding to each key point is extracted from the sample palm vein image. The sample palm vein image is sample data used to train the local feature extraction model. The first sample image block is input into the encoder to obtain the first sample feature vector output by the encoder. The pixel coordinates of the key points are added to the first sample feature vector of the corresponding first sample image block to obtain the second sample feature vector. The second sample feature vector is used as the third feature vector of the second sample image block where the corresponding key point is located, and the zero vector is used as the third sample feature vector of the remaining second sample image block. The second sample image block is obtained by uniformly dividing the sample palm vein image. Each of the third sample feature vectors is input into the decoder to obtain the decoded image output by the decoder; Based on the reconstruction loss function, a first loss value is determined using the decoded image and the sample palm vein image, and the network parameters of the decoder and the encoder are optimized using the first loss value.

2. The method according to claim 1, characterized in that, The local feature extraction model is an encoder; correspondingly, the step of extracting the first local feature vector of each of the first image patches using the pre-trained local feature extraction model includes: Each of the first image blocks is input into the encoder to obtain the first feature vector output by the encoder; The pixel coordinates of the target key points are added to the first feature vector of the corresponding first image block to obtain the first local feature vector of the first image block.

3. The method according to claim 1, characterized in that, When determining a first loss value based on the reconstruction loss function using the decoded image and the sample palm vein image, and optimizing the network parameters of the decoder and the encoder using the first loss value, the method further includes: Based on the nearest neighbor metric loss function, a second loss value is determined using the second sample feature vector, and the network parameters of the encoder are optimized using the second loss value.

4. The method according to claim 3, characterized in that, The step of determining the second loss value based on the nearest neighbor metric loss function using the second sample feature vector includes: Obtain a third palm vein image belonging to the same user as the sample palm vein image, perform key point detection on the third palm vein image, and extract the third image block corresponding to each key point from the third palm vein image based on the key points in the third palm vein image; The third image block is input into the encoder to obtain the second feature vector output by the encoder. The pixel coordinates of the key points are added to the second feature vector of the corresponding third image block to obtain the third feature vector. Based on the feature distance between the second sample feature vector and each of the third feature vectors, the third feature vector that is closest to the second sample feature vector is taken as a feature vector of the same class, and the third feature vector that is second closest to the second sample feature vector is taken as a feature vector of a different class. Based on the nearest neighbor metric loss function, the second loss value is determined using each of the second feature vectors and the corresponding similar and dissimilar feature vectors.

5. The method according to claim 1, characterized in that, The step of determining the target palm vein image from the plurality of candidate palm vein images using each of the first local feature vectors includes: Each of the first local feature vectors is matched with each of the second local feature vectors of the candidate palm vein image; If the matching degree between the first local feature vector and the second local feature vector is greater than or equal to a preset matching threshold, it is determined that the key point corresponding to the first local feature vector matches the key point of the second local feature vector. The candidate palm vein image with the highest number of key point matches is selected as the target palm vein image.

6. A palm vein recognition device, characterized in that, include: The global feature extraction module is configured to input the first palm vein image to be identified into a pre-trained global feature extraction model to obtain a global feature image output by the global feature extraction model. An image block cropping module is configured to perform key point detection on the first palm vein image and, based on the key points in the first palm vein image, crop a first image block corresponding to each key point from the first palm vein image; wherein, the image block cropping module includes: a first image division unit configured to divide the first palm vein image into multiple second image blocks of a first preset size; a first key point selection unit configured to randomly select a first preset number of target key points from each key point in the first palm vein image, wherein each target key point is located in a different second image block; and a first image block cropping unit configured to crop a first image block of a first preset size centered on the target key point from the first palm vein image; The local feature extraction module is configured to extract the first local feature vector of each of the first image patches using a pre-trained local feature extraction model. The identity recognition module is configured to match multiple candidate palm vein images from multiple second palm vein images using the global feature image, determine a target palm vein image from the multiple candidate palm vein images using each of the first local feature vectors, and determine the identity information of the target palm vein image as the identity information of the first palm vein image. The palm vein recognition device also includes a model training module, which includes: The second image patching unit is configured to perform key point detection on the sample palm vein image and, based on the key points in the sample palm vein image, to extract the first sample image patch corresponding to each key point from the sample palm vein image. The sample palm vein image is sample data used to train the local feature extraction model. The second vector extraction unit is configured to input the first sample image block into the encoder to obtain the first sample feature vector output by the encoder, and add the pixel coordinates of the key points into the first sample feature vector of the corresponding first sample image block to obtain the second sample feature vector. The third vector generation unit is configured to use the second sample feature vector as the third feature vector of the second sample image block where the corresponding key point is located, and use the zero vector as the third sample feature vector of the remaining second sample image block. The second sample image block is obtained by uniformly dividing the sample palm vein image. The decoding unit is configured to input the feature vectors of each third sample into the decoder to obtain the decoded image output by the decoder; The first optimization unit is configured to determine a first loss value based on the reconstruction loss function by decoding the image and the sample palm vein image, and to optimize the network parameters of the decoder and encoder based on the first loss value.

7. A palm vein recognition device, characterized in that, include: Memory and one or more processors; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the palm vein recognition method as described in any one of claims 1-5.

8. A storage medium containing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a computer processor, are used to perform the palm vein recognition method as described in any one of claims 1-5.

Images