Image Segmentation Based On Active Contour Models

Image segmentation is a key in the research field of image processing and computer vision, and gets more and more attention in the research and application of image processing technology. Color image segmentation can be regarded as an extension of gray image segmentation methods, however, some of the methods can not be applied to color image segmentation directly.C-V model (Chan-Vese model) is a classical active contour model based on the M-S model, and uses level set method to minimize energy function. C-V model utilizes region information of images so that it has a strong anti-noise ability to segment images with fuzzy boundaries. Although a color image can be segmented by converting to a gray image, the segmentation result is our desired. In order to overcome this difficulty, researchers proposed Multi-Channel C-V model to segment color images.Both C-V model and Multi-Channel C-V model are based on assumption that inside and outside of the evolving curve have consistent feature properties respectively, this is one reason that they can not work well in heterogeneous images. To solve this problem, in this paper we propose a novel active contour model based on superpixels, and also we introduce edge instruction function to get accurate boundaries of objects. In addition we extract color information of images using color index in 8-HSV space. Practical experiments prove that our algorithm can obtain satisfactory segmented results.Another reason that both C-V model and Multi-Channel C-V model can not work well in heterogeneous images, is that they just utilize global statistics of image. Existing localizing region-based active contour method can deal with this trouble, however, the method obtain incorrect result with contour trapped in homogeneous regions. In this paper we propose a novel active contour model based on global and local statistics. First, find and update the points which are trapped in homogeneous regions. Then introduce edge instruction function to and global information to restrain the contour staying near real boundaries of objects. Finally, we put the novel method into the framework of co-segmentation. Practical experiments prove that our algorithm can obtain satisfactory segmented results.