Multiple Distribution Active Contour For Image Segmentation

The active contour model is a main method in the field of image segmentation.This model has effective segmentation, matching and tracking analysis with the priorknowledge of the image. However, because of the non-convexity structure, thesegmentation results generally easy to fall into the predicament of the local solution.Thus the result of some image segmentation is not very good.To solve the problem of active contour model being easy to fall into local solution,scholars therefore, have proposed many global algorithm combined with convexoptimization methods. Then a number of global algorithms are combined to the imagesegmentation, such as the fast global minimization of active contour snake model,local histogram based on Wasserstein distance and fuzzy region competition usingspatial/frequency information, which partly achieve ideal segmentation. But all thesealgorithms are also facing with the parameter self-adaptive problems, namely they allneed to artificially adjust some parameters, which greatly reduces the efficiency.In order to solve the problem of local solution and the parameter adaptive at thesame time, an active contour model is proposed for image segmentation based onmultiple distributions on each region with adaptive weight between them in this paper.The distributions are drawn from different filtering results of original gray image in theregion, referred as projections under different filters. Because in image segmentation,the different characteristics of different contribution to segmentation is difficult todistinguish, in this paper, all the characteristics of the information are expressed by theprojection distributions, which determine the influence of different characteristics onthe energy functional.The energy associated to a region is described as the weighted sum of the entropyof all projection distributions on it. The weight coefficient for each distribution isdecided by the ratio of its entropy compared to others in the region. The segmentationprocess is to minimize the sum of weighted entropy for all regions, in which throughthe adaptive weight coefficient the competition between different distributions on eachregion is embodied. The minimization of the proposed energy model is achieved withlevel set method. Through experiments contrast with other methods, the results show that the proposed method can achieve automatic segmentation of many natural imageswith minimal parameter.