Effectively Segmentation Of Natural Image By Using Lossy Compression

In the field of computer vision, image segmentation (Segmentation)refers to the digital image into multiple image sub-regions (pixel sets,also known as super-pixels) in the process. The purpose of imagesegmentation is to simplify or change the representation of images,making images easier to understand and analyze. Image segmentationis usually used to locate objects in the image and boundaries (lines,curves, etc.). More accurate, image segmentation is an image of eachpixel tagging a process that makes the pixels with the same label with acommon visual feature. Currently there have been thousands of imagesegmentation methods, but not an image segmentation method has avery good segmentation result for all the images. Image segmentationis a critical step of image processing to image analysis, the pros andcons of the segmentation results will directly affect subsequent imageanalysis, image understanding and scene recovery. Therefore, imagesegmentation has its importance and relevance.In this paper, we cast natural-image segmentation as a problem ofclustering texture features as multivariate mixed data. We model thedistribution of the texture features using a mixture of Gaussiandistributions. However, unlike most existing clustering methods, we allowthe mixture components to be degenerate or nearly-degenerate. Wecontend that this assumption is particularly important for imagesegmentation, where degeneracy is typically introduced by using acommon feature representation for different textures. Our goal is to findthe optimal segmentation that minimizes the overall coding length ofthe segmented data, subject to a given distortion. By analyzing thecoding length/rate of mixed data, we formally establish some strongconnections of data segmentation to many fundamental concepts inlossy data compression and rate-distortion theory. We show that a deterministic segmentation is approximately the (asymptotically)optimal solution for compressing mixed data. We propose a very simpleand effective algorithm that depends on a single parameter, theallowable distortion. At any given distortion, the algorithm automaticallydetermines the corresponding number and dimension of the groupsand does not involve any parameter estimation. Simulation resultsreveal intriguing phase-transition-like behaviors of the number ofsegments when changing the level of distortion or the amount ofoutliers. We show that such a mixture distribution can be effectivelysegmented by a simple agglomerative clustering algorithm derivedfrom a loss data compression approach. Using simple fixed-sizeGaussian windows as texture features, the algorithm segments animage by minimizing the overall coding length of all the feature vectors.In our comprehensive experiments, we compare the segmentationresults of the proposed algorithm in this paper with commonly usedCanny edge detection algorithm, the gradient algorithm of Sobeloperator, K-means algorithm and fuzzy C-means algorithm, andmeasure the performance of the proposed algorithm by adjusting thetexture threshold λ to optimize the segmentation results. Theseexperimental results show that the segmented image by using ouralgorithm is closer to human subjective judgmental segmentation.