Active contour model based on local absolute difference energy and fractional-order penalty term

Abstract

The active contour model (ACM) is one of the popular image segmentation methods. In practice, however, images are often corrupted by severe noise and intensity inhomogeneity, making traditional ACMs unavailable. In this paper, a novel ACM is presented to effectively cope with noise and intensity inhomogeneity. To better discriminate the object and background regions in inhomogeneous images, a new saliency fitting image (SFI), defined as the difference between the original image and the local-weighted image (LWI), is first constructed. In addition, the local mean absolute deviation (LMAD) based weight factor can reflect the degree of the intensity variation in a local region, and the LWI represents the bias component accounting for intensity inhomogeneity. Then, a local fitting energy term is defined as the absolute difference between two local fitting functions, which represent the local averages of image intensity inside/outside the contour in two windows with different sizes. And a threshold strategy is utilized to improve the robustness of the proposed model to noise. Finally, a fractional-order penalty term is defined to penalize the deviation between the level set function (LSF) and the signed distance function (SDF), which makes the proposed model avoid the side effect of the penalty energy term in traditional ACMs. Experimental results show the efficiency and accuracy of the proposed ACM.