| With the increased demand for reliability and accuracy indicators of biometrics,finger vein recognition has become an important branch of extensive research.Compared with other biometrics,finger vein recognition comes with live recognition,and is relatively simple to apply,so it has important research significance.In recent years,with the wide application of deep learning,the accuracy and robustness of recognition algorithms have been continuously improved,but there are still challenges in how to further improve robustness.The key research and innovation of the thesis are as follows:(1)In order to reduce the impact of image perturbation on the robust features extracted by algorithm,supervised and unsupervised data enhancement methods are adopted,supplemented by normalization processing,which enriches the amount of data and reduces overfitting.In order to solve the problems of fog blur and uneven illumination of finger vein image quality,the method of combining dehazing and histogram equalization with contrast limited is adopted.To reduce the impact of image backgrounds,the deep segmentation network R-LAB(Refine-Laboratory)is designed and implemented.Add upsampling branches to the shallow downsampling process and introduce skip connection to improve the utilization of shallow texture feature;Replace traditional pooling with separable atrous group convolution to reduce information loss;At the junction of upsampling and downsampling,the atrous spatial pyramid pooling is connected to obtain multi-scale deep semantic features;Focal loss is used to reduce the impact of sample pixel imbalance.(2)Aiming at the low prediction accuracy of the existing finger vein image classification network,a robust classification network with modified Xception is proposed.After every convolutional block structure of the original Xception network,the global average pooling process is added,and multi-scale feature prediction is carried out,and the fusion pooling characteristics are used as the final classification basis.Add a residual compensation structure on the residual branch to prevent the omission of shallow information;Enhance generalization performance with transfer learning to accelerate model training;Design cross-entropy loss with label smoothing to mitigate overfitting.Experiments show that the accuracy of classification prediction on the self-built dataset SF＿DA(Self＿Dataset)is 99.78%,and the accuracy of the classification prediction on the public dataset SDU-FV is 99.84%.(3)This paper proposes a recognition algorithm based on network feature output combined with KNN classifier,which realizes feature extraction with weighted classification model and uses cosine distance as the classification metric,which can realize the recognition of unregistered samples without training new data.In order to improve the robustness of the classifier,the double discriminant threshold and the two-segment discriminant interval are set,and the contingency of single interval recognition is reduced in a priority-weighted manner.Experiments show that the recognition accuracy on the self-built dataset SF＿DA is 99.11%,and the accuracy loss is 0.67%.The recognition accuracy on the public dataset SDU-FV is 99.53%,and the accuracy loss is 0.31%.This paper integrates the end-to-end and KNN recognition methods into an integrated framework,including registration,collection,login,recognition,database query and management and other functions. |
