Local information enhanced LBP

ZHANG Gang(张刚)^{1, 2}, SU Guang-da(苏光大)¹, CHEN Jian-sheng(陈健生)¹, WANG Jing(王晶)¹

(1. Department of Electronic Engineering, Tsinghua University, Beijing 100084, China;
2. School of Software, Shenyang University of Technology, Shenyang 110023, China)

Abstract:Based on the observation that there exists multiple information in a pixel neighbor, such as distance sum and gray difference sum, local information enhanced LBP (local binary pattern) approach, i.e. LE-LBP, is presented. Geometric information of the pixel neighborhood is used to compute minimum distance sum. Gray variation information is used to compute gray difference sum. Then, both the minimum distance sum and the gray difference sum are used to build a feature space. Feature spectrum of the image is computed on the feature space. Histogram computed from the feature spectrum is used to characterize the image. Compared with LBP, rotation invariant LBP, uniform LBP and LBP with local contrast, it is found that the feature spectrum image from LE-LBP contains more details, however, the feature vector is more discriminative. The retrieval precision of the system using LE-LBP is 91.8% when recall is 10% for bus images.

Key words:texture feature extraction; LE-LBP; minimum distance sum; gray difference sum