Democratic voting downsampling for coding-based palmprint recognition

Abstract

Downsampling is critical for coding-based methods to reduce storage and accelerate matching speed. In coding-based palmprint recognition methods, the image size of the region of interest is typically 128 × 128, which is divided into 32 × 32 blocks and each block consists of 4 × 4 pixels. In the traditional downsampling method, the upper-left pixel in each block is selected to represent the feature of this block. However, this crude technique likely leads to serious information loss. The feature template heavily depends on the upper-left pixels, which degrades the tolerance for pixel-level dislocation and rotation. The authors analyse the downsampling stage in depth, and propose a democratic voting downsampling method (DVDM), which can improve the robustness and accuracy of the coding-based palmprint recognition methods without any prior knowledge. All the pixels in each block have equal voting rights to determine the feature of this block, so DVDM can extract stable features and effectively overcome the autocracy of an upper-left pixel. The sufficient experiments tested on the public PolyU palmprint dataset to confirm that DVDM can remarkably improve the robustness to pixel-level dislocation and rotation, and also improve accuracy performance, equal error rates of the coding-based methods are dropped down at most 11.5%.