WIRIF: Wave interference-based rotation invariant feature for texture description

Abstract

Abstract This paper presents a novel wave-interference based rotation-invariant feature for texture description. In local pattern-based techniques, features are extracted from small circularly symmetric neighboring points. These features require further processing to achieve rotation invariant property. In addition, small neighborhoods can only capture the information of small areas which is suitable for describing micro-textures. The proposed method in this paper is based on wave-interference signals generated at each pixel of an image that needs no more operations for achieving rotation invariance. We show that the wave-interference signals are inherently rotation invariant in addition to their ability in adaptively combining information of larger asymmetric areas with adaptive weights. We use discrete version of wave equation to compute wave function in 2D plane of image. For extracting final features, we employ marginal histograms of sampled interference signals to extract rotation-invariant features. We also improved the discrimination power of the proposed features by including several orders of differences of sampled interference signal. The experimental results on Outex, Brodatz, VisTex, and KTH-TIPS2 datasets show that the proposed method outperforms the prior techniques regarding precision and recall in texture retrieval application.