Research Of Remote Sensing Image Segmentation Based On An Enhanced Multiphase Level Set Method

Image segmentation is an important issue in computer vision and patternrecognition, and also the most important aspect in Image Engineering. The purpose ofimage segmentation is to divide the image to a certain number of sub-regions whichshare the same properties, only after the segmentation can meet the needs of imageanalysis. In recent years, more and more methods have been developed for imagesegmentation by researches, among these methods, the algorithm that based on PartialDifferential Equations (PDE), is capable of handling the topological change in curvesand besides it has a strict mathematical theory basis which offers a good stability andhigh accuracy of the algorithm, because of these advantages, it has been widelyapplied in image segmentation.The active contour model is of the representation method of PDE based methodsand it can be divided into two categories: one is Edge-based model; the other isRegion-based model. The Edge-based model is mainly utilizing the information onthe edges of the targets, such as curvature, gradient, etc., but there exist somelimitations, in the case of the edges of targets is blurred or broken and the case oftopological changes in curve evolution, the edge based model cannot handle theseproblems. Region-based model can solve the aforementioned problems because of theusing of global information, such as intensity, shape, color, etc. Especially theemergence of the level set method, it implicitly express the evolution curve of zerolevel set function, thus avoiding the evolution tracking and processing parameters,providing a precise and stable mathematical model, successfully solved the situationof split and merge on the evolution curve, so that more and more attention is paid tothe Region-based model.Remote sensing images with the characters of multi-gray-level, multi-target,complex terrain information, blurred boundaries, etc. and using the Edge-based modelto segment these images, usually can’t get satisfied results. So this paper chose the Region-based model as our research object. More specifically, the multiphase C-Vmodel was enhanced to achieve better results. Research contents and innovations ofthis paper are as follows:(1) Firstly, the initial contour of traditional C-V model can be anywhere on theimage, but actually we find that different initial contours play an important role insegmentation result. If not properly chosen, on the one hand, it will lower the speed ofthe convergence; on the other hand, it may cause the failure of the segmentation. Inorder to solve this, we introduce a new initial method which is to extract the contoursof the first n (n is the number of level set functions) biggest area as our initialcontours, experimental results proves that the new method can speed up theconvergence and make the segmentation more stable.(2)Secondly, traditional C-V model needs re-initialization to keep the level setfunction to SDF function during the evolution process, which is very time-consuming.To solve this, we add the distance penalty term to the C-V model to eliminate theproblem.(3)At last, using traditional C-V model to segment remote sensing image, there stillexist some trivial regions on the result which is redundant. Hence we add the gradientinformation, together with the global intensity information, to boost the evolution andeliminate the trivial regions.