Hand vein recognition based on multi-scale LBP and wavelet

This paper proposes a novel approach to extract multi-scale LBP features of hand vein images. Local Binary Patterns (LBP) is a simple but efficient texture operator for hand vein images. However there are two big limitations of LBP. One is that it only covers a small neighborhood area and gets very limited local information. The other is some non-uniform information will be lost as uniform patterns is adopted. To solve these problems, the hand vein image is decomposed with two-level wavelet, and gets 8 coefficient matrices: A₁, H₁, V₁, D₁, A₂, H₂, V₂ and D₂. In order to obtain the weight of each sub-band, the recognition accuracy of each sub-band is calculated by original LBP based on Euclidean Distance. In consideration of the direction of hand vein images, two diagonal high frequency components, D₁ and D₂, are abandoned. Meanwhile, A₁, H₁, V₁, A₂, H₂, V₂ and the original image are chosen as multi-scale components. Finally all LBP features of components are fused. Experimental results on a database of 2040 images show that the proposed method outperforms original LBP and traditional multi-scale LBP.