Finger Vein Recognition Based on Multi-Task Learning

Abstract

In finger vein recognition, traditional methods for extracting ROI based on edge detection, sliding window detection of joint lines, etc. need to set a fixed threshold, which contains many parameters that need to be adjusted. In the case of large illumination changes or poor image quality, the extracted results are not accurate enough. The existing feature extraction method also has a fixed operator pattern and limited extracted feature patterns. Therefore, a large amount of effective feature information is wasted. In this paper, a multi-task neural network model algorithm is proposed, which uses the multi-task learning method to jointly optimize the ROI extraction task and the feature extraction task. This method not only improves the overall data processing efficiency of finger vein recognition system, but also improves the quality of extracted vein features. At the same time, we explore the use of improved loss function based on softmax to train our model. Our model is better than traditional methods and single task neural network model algorithm in MMCBNU [16] FV-USM [17] and DUMLA-HMT [18] data sets.