Edge-preserving Image Smoothing Alogrithm

Plenty of natural photos and art pieces embodies complex texture and well-structured objects.Edge-preserving image smoothing aims to eliminate the detail element in the input image under the circumstance that initial edges are carefully preserved.However,the complicacy of detail patterns,in many cases,hinders many current algorithms from producing desired results.Some edge-preserving smoothing methods regard high-contrast pixels as edges,which protects large-amplitude details instead of smoothing them.Some approaches attempt to keep large gradients of pixels rather than intensities,which deviates the final results from their initial shape.We generalize the drawbacks of these methods as: detail residue,edge collapse,and shape distortion.This paper introduces a new way to generate edge-preserving image smoothing outcome with properties of detail-smoothing,edge-fitting,and shape-retaining.This strategy consists of three major steps: exploiting the windowed variation similarity measure to suppress highcontrast details,extracting edge information by detecting salient leap,and eventually imposing the edge-guided image to the weighted least squares framework to recover the initial shape.The detail suppression phase represses the gradients of high-contrast fine-scale detail to lower magnitudes than those of edges,so that details and edges are easy to distinguish.The edge extraction phase collects the possibility information of a pixel pertaining to an edge,so that the edges of the input image can be preserved.The shape recovery phase rectifies the overly-smoothing and distortion problem raised in the previous phases,which is significantly important,since many scenarios require delicate manipulations on details.Our approach can not only tell edges and details in similar scale apart,but also preserve the original structure of input image.Edgepreserving image smoothing technique is the foundation of many computer vision applications.We carry out some applications,such as detail enhancement,edge extraction,image abstraction,and HDR tone mapping,to demonstrate the effectiveness and robustness of the algorithm in this paper.