| With the rapid development and widespread application of Internet technology,information security has received increasing attention.Palm vein recognition technology,as a new identity authentication technology,has been widely studied for its advantages of high precision,high security,and live recognition.However,due to various factors such as lighting,light scattering,and pose changes,the quality of captured palm vein images is not ideal,which reduces the system’s recognition performance.In addition,existing palm vein classifiers use logical labels(one-hot)for training.However,logical labeling is based on the assumption that the labels of all sample data are independent of each other.In the actual collected vein library,especially for data with high similarity,the labels implicitly contained in the data are related.In order to solve these problems,this thesis applies deep learning to the extraction and recognition of palm vein features and conducts the following research work.(1)Proposed a Label-Enhanced Multi-Scale Vision Transformer(LE-MSVT)model to achieve effective palm vein recognition.Firstly,a Multi-Scale Vision Transformer(MSVT)model is established to learn the multi-scale feature of palm vein images.The model consists of convolutional and self-attention modules,where the convolutional module is designed to learn local feature information of an image,and the self-attention module mainly learns the dependency relationship between images of different scales.Secondly,a Label-Enhanced Graph Convolutional Network(GCNLE)model is constructed to learn the true label distribution of the samples and improve the recognition performance of the vein classifier.The model uses a multi-layer perceptron to learn a label-related matrix to capture the relationship information between the input image and each training image in each category.Then,the label vector generated by GCN is combined with the logical label(one-hot)vector to calculate the true label distribution of the input image.Finally,the GCNLE model is applied to the MSVT classifier to establish the LE-MSVT model.The accuracy of the model was 99.70%,95.45% and 98.73% on the three public database sub-datasets.(2)A label-enhanced hybrid self-attention palm vein recognition method is proposed.First,a hybrid self-attention module(RMT)is established,which includes self-attention module,channel attention module,and spatial attention module.Then,multiple RMT modules are stacked to construct a Transformer model that integrates multi-path attention to achieve palm vein image recognition.Because the hybrid self-attention module integrates global attention,channel attention,and spatial attention features,the model can learn robust vein features.Secondly,a label-enhanced model based on attention mechanism is established(AttentionLE).The model learns the correlation between samples through the self-attention mechanism to generate the true label distribution of the input samples.Finally,the label-enhanced hybrid self-attention palm vein recognition model is obtained by combining the AttentionLE and RMT modules.The model achieved 99.77%,97.15%and 96.82% accuracy in three public databases. |
