基于级联U-Net的手指静脉分割算法研究

Research on finger vein segmentation algorithm based on cascaded U⁃Net

针对传统的手指静脉分割算法无法达到良好的分割效果,公开手指静脉数据集较小,没有合适的参考标准用于神经网络训练,数据集未进行ROI处理导致包含背景过多等问题,提出一种基于级联U-Net的手指静脉分割算法。从UNet网络结构出发,针对所使用的指静脉数据集简化网络、减少参数,并以此作为所提方法的网络组成部分。针对山东大学公开数据集SDU-FV采用一种基于最小二乘的旋转校正算法提取ROI区域,该方法可以克服手指轴向旋转引起的成像区域不一致等问题,进而突出静脉丰富区域。在神经网络训练过程中,一方面在每张图像中随机选择其中心获得子块进行数据扩充;另一方面将检测静脉图像横截面局部最大曲率方法提取到的纹路作为标准。实验结果表明,基于级联U-Net的手指静脉分割算法取得了比其他方法更加优异的分割性能,在3个手指静脉公开数据集SDU-FV、MMCBNU_6000和THU-FVFDT2上分别获得了91.56%,92.91%和93.08%的AUC以及92.44%,93.93%和94.86%的准确率。

A finger vein segmentation algorithm based on cascaded U⁃Net is proposed to cope with such problems as the traditional finger vein segmentation algorithm fails to achieve a good segmentation effect,the finger vein public datasets are not large enough,reference standard suitable for neural network training is unavailable,and datasets are not subjected to ROI processing,which results in redundant backgrounds.On the basis of the U⁃Net network structure and finger vein dataset,the network is simplified and the parameters are reduced.It is taken as a network constituent part of the proposed method.According to the public dataset SDU⁃FV of Shandong University,a rotation correction algorithm based on least squares is used to extract the ROI region.This method can be used to overcome the problem of inconsistent imaging area caused by the axial rotation of the finger,and then highlight the vein⁃rich region.In the process of neural network training,the center of each image is randomly selected to obtain sub⁃blocks for data augmentation on the one hand;on the other hand,the patterns extracted with the method of detecting local maximum curvature in cross⁃section of a vein image are taken as the standard.The experimental results show that the finger vein segmentation algorithm based on cascaded U⁃Net obtains an AUC(area under the curve)of 91.56%,92.91 and 93.08%and accuracy of 92.44%,93.93%and 94.86%from 3 finger vein public datasets of SDU⁃FV,MMCBNU_6000 and THU⁃FVFDT2 respectively,whose segmentation performance is more superior than that of other methods.