Siamese-Hashing Network for Few-Shot Palmprint Recognition

Abstract

In recent years, palmprint-based recognition technology has become one of the hotspots in biometrics research. The accuracy of traditional palmprint recognition algorithms mainly depends on vast data and labels. However, in reality, we usually have few labeled data. To solve this problem, the paper explores the application of few-shot recognition to palmprint. In the preprocessing stage, a novel region of interest (ROI) extraction algorithm is proposed, which can extract more palmprint texture features in the relatively fixed palm area and effectively improve the impact of palm size on preprocessing results. In the feature extraction stage, the paper presents a nonpooling Siamese-Hashing Network structure, called SHN. This method can extract high discriminant features of new categories from only a small number of samples. In addition, the output of SHN is a 48-bit hashing code, which takes up less memory and matches samples faster. Experiment results show that the performance of the model in the benchmark database is better than other classical models in the few-shot case.