LAWNet: A Lightweight Attention-Based Deep Learning Model for Wrist Vein Verification in Smartphones Using RGB Images

The wrist vein is one robust and reliable biometric for research and applications in automatic human verification; however, the existing wrist vein recognition models were heavy and ineffective in deploying on smartphones. Also, smartphones were required to integrate near-infrared (NIR) sensors to capture wrist vein images. This article proposes a novel lightweight attention-based deep learning model for wrist vein verification in regular smartphones (LAWNet) using red-green-blue (RGB) images. We first used a saturation channel converted from an RGB image as LAWNet’s input rather than a NIR image, as in previous research. Second, we presented a new wrist region of interest (ROI) extraction method. This step was usually neglected in previous research on contact wrist NIR databases; however, a correct wrist ROI can improve the final recognition performance in a contactless environment. Third, we proposed a novel LAWNet model for wrist vein extractors. Our model was formed by one convolutional block, four improved sandglass blocks, followed by a mixed pooling coordinate attention (MPCA) module, and completed with one parallel spatial pyramid pooling (PSPP) block. The experiment results showed that the proposed model has fewer parameters but achieved the best performance among compared models on our self-collected RGB wrist vein database with an equal error rate (EER) of 1.80%; furthermore, the speed of the complete wrist vein recognition system deployed on smartphones is extremely fast at 60 frames/s, which is acceptable in practice.