Axially Enhanced Local Attention Network for Finger Vein Recognition

Yiwei Huang,Hui Ma,Mingyang Wang

DOI: https://doi.org/10.1109/tim.2023.3291785

IF: 5.6

2023-07-20

IEEE Transactions on Instrumentation and Measurement

Abstract:Compared to other biometric features, finger vein depicts vascular topology including some long-distance vascular and local textures and has received much attention for their advantages in liveness detection and convenience of collection. However, some recent studies have focused on global-based spatial attention mechanisms to enhance the capture of long-distance information but ignore local texture information, which is usually computationally inefficient. To address this problem, we propose an axially enhanced local attention network (ALA Net) that uses the distribution of vascular topology to infer attention. Specifically, the proposed ALA locates the receptive domain of each representation to its deformable neighborhood, thus extracting the vein pattern more flexibly and accurately. In addition, we propose a feature amplification strategy using a low-cost group convolution to improve the feature similarity and characterization accuracy between channels, resulting in higher accuracy at lower parameter scales. The effectiveness of the method is illustrated by searching the acceptance domain of the ALA module using image distribution with the accuracy of 94.50% on the SDUMLA dataset, 97.86% on the HKPU dataset, and 98.92% on the NUPT dataset, and the advancement of the ALA Net is demonstrated by comparing it with the state-of-the-art attention model and backbone network.

engineering, electrical & electronic,instruments & instrumentation