Extreme Downsampling and Joint Feature for Coding-Based Palmprint Recognition

Abstract

Palmprint is a commonly used and significant modality for biometric recognition currently. The coding-based methods are practical for palmprint recognition because they can be free from training, and have low computational complexity and storage cost. Downsampling is widely used in encoding-based methods for storage cost reduction and matching speed acceleration. In the traditional downsampling method (TDM), the feature of each block is only determined by its upper left pixel; however, TDM totally ignores the other useful pixels. We propose a practical but effective general downsampling method, dubbed extreme downsampling method (EDM). In EDM, the best response pixel in each block is selected as the representative. Because the winner is selected from the 16 candidates in each block, EDM substantially improves the robustness against the dislocation due to imperfect preprocessing, and accordingly improves the accuracy performance. In addition, we propose a joint feature method (JFM) to fuse the matching distances of the templates at score level, which are downsampled by TDM and EDM. The accuracy performance and robustness of EDM and JFM are solidly confirmed when they are employed on several state-of-the-art coding-based palmprint recognition methods.