Performance Improvement of Dorsal Hand Recognition via Multi-band Selection

Abstract

A great advantage of multispectral technique is to pursue better recognition performance through band fusion. Adding more bands can build larger feature dimension space while bringing in more redundant information. This paper tries to optimize band set for multispectral dorsal hand recognition mainly in near infrared (NIR) light. Images at 35 bands are sampled uniformly from 700nm to 1040nm, and then band distribution is analyzed for multispectral biometric specialty. Multi-band selection is processed in two steps. First, the whole NIR region is divided into several band clusters according to maximum irrelevance principle. Second, representative bands are chosen from these clusters for recognition rate ranking. Our scheme focuses on accuracy and rapidity simultaneously. Experiment shows the robustness of improved clustering method and the high fusion performance. The proposed band selection method can be applied to other multispectral database with consecutive band feature change.