Real-Time Segmentation Method of Lightweight Network For Finger Vein Using Embedded Terminal Technique

Abstract

Because the existing finger vein segmentation networks are too large and not suitable for implementation in mobile terminals, the reduction of the parameters of the lightweight network leads to the reduction of the segmentation index, and the long-running time of deep network on hardware platforms; this paper proposes a lightweight real-time segmentation method for finger veins based on embedded terminal technique. In the preprocessing stage of the algorithm, the data is greatly expanded by randomly selecting the center to obtain sub-blocks on each image of the training set. The network first uses deep separable convolution to greatly reduce the U-Net parameters of a basic network and introduces an attention module to reorder the features to improve network performance, followed by a preliminary lightweight network Dinty-NetV1. Second, the Ghost module is added to the deep separable convolution, and the feature map of the network part is obtained through a cheap operation so that the network is further compressed to obtain Dinty-NetV2. After adding channel shuffle, all the characteristic channels are evenly shuffled and reorganized to obtain Dinty-NetV3. Finally, a study of the filter norm yields the distribution characteristics of the finger vein picture features. By using the geometric median pruning method, the network models for each stage of the algorithm proposed in this paper achieved better segmentation performance and shorter split time after pruning. The overall Dinty-NetV3 model size is only less than 9% of the U-Net and Mult-Adds is less than 2% of the U-Net with the same structure. After testing on two-finger vein datasets SDU-FV and MMCUBV-6000, we confirm that the performance of Dinty-NetV3 surpasses all previously proposed classic compression model algorithms and it is not inferior to more complex and huge networks such as U-Net, DU-Net, and R2U-Net. The proposed algorithm has advantages in terms of time needed to train the network, and we verify its universality using NVIDIA’s full range of embedded terminals.