A finger vein feature extraction network fusing global/local features and its lightweight network.

Finger vein recognition technology has been widely used in various identity authentication scenarios due to its convenience, fast recognition speed, and high security. However, there is an inevitable problem that seriously affects its recognition, which is the change of finger poses, such as shift and rotation during the image acquisition process. Therefore, how to extract finger vein features that are more robust to finger pose changes is a more concerned issue. We analyze the multi-pose finger vein images in practical application scenarios, and find the fact that the global vein features of the same finger are quite different while the local features are highly similar. In most existing finger vein recognition algorithms based on convolutional neural network (CNN), only global finger vein features are extracted by global average pooling (GAP), which fails to take full account of the above fact. In this paper, we proposed a finger vein feature extraction network fusing global and local features (FGL-Net) and its lightweight network KD-FGL-MobileNet, which effectively improves finger vein recognition performance in different finger poses. Firstly, FGL-Net consists of two parts, the backbone network based on the designed ResBlock with Mish is used to extract the high-level semantic features of finger veins, and then the global and local feature extraction module with three independent branches are designed to fully learn the global and local finger vein features at different granularities, and finally all the features of the three branches are fused into a fusion feature with greater robustness to pose change for recognition. To improve the generalization ability of the network, the CurricularFace loss is added to train FGL-Net with the cross-entropy loss. Such design not only aggregates homologous features and separates heterologous features, but also mines finger vein image samples under special poses online for intensive training. Secondly, according to the characteristics of the finger veins, we design a lightweight residual block based on fast receptive field (SE-FrfResBlock) to build a more lightweight FGL-MobileNet. A knowledge distillation loss and a feature map loss are added to FGL-MobileNet to address the generalization performance degradation of FGL-MobileNet, and we named it KD-FGL-MobileNet. On FV-USM, FV-Normal and FV-Specical datasets, compared with VGG-Net and InceptionResnet, the Top1 ranking of FGL-Net are improved by 2.38%, 8.26%, 13.42% and 0.00%, 1.17%, 9.95%, and the recognition rate are improved by 9.42%, 18.59%, 28.33% and 0.72%, 7.12%, 20.88%. On FV-Specical datasets, compared with MobileNetv3, FGL-MobileNet, the recognition rate and Top1 ranking of KD-FGL-MobileNet are improved by 24.12% ,6.19% and 12.46%, 3.78%. The above results show that the proposed FGL-Net effectively improves the recognition performance of finger vein images in different poses, and KD-FGL-MobileNe requires less storage space while remaining basically consistent with the performance of FGL-Net. [ABSTRACT FROM AUTHOR]