Structure-Preserving Image Smoothing via Phase Congruency-aware Weighted Least Square

Abstract

Structure-preserving image smoothing, or also understood as structure-texture separation problem, is an important topic for both computer vision and computer graphics as structure-texture separation can help better image understanding. In fact, many image processing problems can be well achieved once two layers possessing different properties of a scene are separated. Therefore better separating structure and texture from an image is of great practical importance. However, it is also a challenge topic since it is often quite subjective to tell the difference between the two layers. Recently, researchers made great efforts on separating a given image into its structure and texture layers by distinguishing edges from oscillations based on non-gradients-based descriptors or descriptors defined specifically for certain kinds of image data. These methods show advantages compared to the purely gradients-based methods with extra information provided besides gradients. In this paper, we propose a structure-texture separation method using non-gradients-based descriptor. Specially, we propose an alternative yet simple image smoothing approach based on the well-known weighted least square (WLS) framework. Our approach combines the phase congruency features that can better help locate structure or contour information of objects. Phase congruency performs well for distinguishing the structure and texture as it mimics the response of the human perception system to contours and is also sensitive to periodic patterns. By including the phase congruency as weights, WLS can better smooth out images while preserving structures. Experimental results indicate that the proposed approach is effective for structure-texture separation and achieves low computational complexity, compared to the state-of-the-art methods.