A Novel Method for Finger Vein Segmentation

In terms of such problems as the traditional finger vein segmentation algorithm cannot achieve good segmentation effect, the public finger vein dataset is small, and there is no suitable reference standard for neural network training, this paper proposes a finger vein segmentation algorithm based on LadderNet. Based on the conventional U-Net structure, we simplify the network and reduce the parameters in view of the characteristic of the finger vein dataset, and U-Net network is taken as part of the LadderNet. By splicing the feature channels of expanding path and contracting path in the network, the semantic information of the image can be obtained as more as possible on the basis of good venous details. With the increase of transmission paths, more complex venous characteristics can be captured. In the process of neural network training, we randomly select the center of each image to obtain sub-blocks for data augmentation; on the other hand, the patterns extracted by detecting the local maximum curvature in cross-sectional of a vein image method are taken as the gold standard, which can extract the centerlines of the veins consistently without being affected by the fluctuations in vein width and brightness, so its pattern matching is highly accurate. We tested this method on two benchmark datasets such as SDU-FV and MMCBNU_6000, the experimental results show that LadderNet-based finger vein segmentation algorithm has achieved superior performance with an AUC of 91.56%, 92.91% and an accuracy of 92.44%, 93.93% respectively over methods in the literature.