Multi-scale local binary patterns based on path integral for texture classification

Abstract

Local binary pattern (LBP) is an effective image texture descriptor due to its high discrimination and easy computation. Moreover, as an extension, multi-scale LBP (MS-LBP) has also been explored for enhancing the conventional LBP and its miscellaneous variants by combining local image structures of different scales. However, since LBPs of different scales are simply combined in a concatenate or joint way, the cross-scale correlation is not fully utilized in MS-LBP. Based on this thought, we propose in this paper a new LBP variant named path integral based LBP (pi-LBP). Specifically, unlike MS-LBP which encodes local patterns individually in each scale, the different scales pixels along a specific path are filtered and then encoded in pi-LBP. In this way, by taking different paths and filters, pi-LBP can effectively encode the cross-scale correlation and provide a better texture description. Experimental results on Outex texture suites show that the proposed pi-LBP outperforms MS-LBP and some other LBP variants on texture classification.