CECNet: Coordinate Encoding Competitive Neural Network For Palm Vein Recognition by Soft Large Margin Centralized Cosine Loss

Abstract

Palm vein recognition plays a crucial role in identity verification, requiring highly discriminative features. However, for touchless palm vein datasets, selecting a suitable and fixed Region of Interest (ROI) is challenging due to variations in capture scales and anisotropy. Moreover, manually annotating ROIs for each palm vein image is a time-consuming and labor-intensive task. To address these challenges, the method based on competitive mechanism is currently a popular approach for palm vein recognition. However, traditional competitive mechanisms only focus on selecting winners from different channels without considering the spatial information of features. In this paper, we reformulate the traditional competition mechanism and propose a Coordinate Encoding Competitive Neural Network (CECNet). Our method takes into account the spatial competition relationship between features, which means we pay attention to features of different directions and scales. We also perform spatial encoding on the competitive features to extract a more comprehensive set of competitive features. To extract the textures, the CECNet employs three parallel Adaptive Gabor Filter Encoders (AGFEs) to learn features of different directions and scales, effectively capturing the variations present in palm vein images. To enhance feature discrimination, the Soft Large Margin Centralized Cosine Loss (SLMCC) Loss function is utilized, taking into account with intra-class similarity and introducing centralized cosine similarity to achieve better interclass separation. By optimizing this loss function, the network learns to prioritize and rank features based on their importance. Experimental results on public palm vein datasets demonstrate the effectiveness of the proposed approach.