Research Of Random Walk For Image Segmentation

Image segmentation is one of the oldest and most widely studied basic problemsin the field of computer vision. It is an important technology of image analysis andkey step of connecting image processing and image understanding and occupies avery important position in image engineering. Up to day, Image segmentation hasproduced a lot of prominent and effective methods. Among these methods, graphapproaches for image segmentation have become popular method because of itsmaturity of the theory and better mapping between image and graph. Random walk isa semi-supervised image segmentation method based on graph theory with fastcomputation、 noise robustness、acceptable responsiveness to weak boundaries andeasy to extend to Multi-dimensional space. However, the performance of randomwalk is susceptible to complex texture image for it adopted gray scale different todescribe the similarity of pairwise nodes and generates incorrect segmentation results.The method builds on graph structure and encounter another deficiency of computespeed will dramatically slows along with the increase nodes and limits its applications.Further study of random walk algorithm has been made in this paper. the mainresearch work includes the following aspects:(1) The computational efficiency is greatly reduced along with the increase ofimage size and pixel. For above problem, a region-based random walk imagesegmentation algorithm was proposed in this paper. First, low level methods wereutilized to generate homogeneous regions which can be used as verticals of graphstructure. Then, the similarity between neighbor nodes was estimated according toregion descriptor and similarity measurement. Finally, random walk algorithm wasrun on the weighed graph and obtained the segmentation result. Compared with thetraditional random walk algorithm, the propose approach could improve thecomputation efficiency while reserve the advantages of traditional methods.(2) Traditional random walk algorithm use gray scale different to describe thesimilarity of neighbor nodes and possess undesirability performance to complextexture image and color image; region based random walk segmentation proposeabove adopted color histogram as region descriptor, which can not sufficiently reactthe spatial structure features. An improved random walk image segmentationcombined with texture features had been put forward to solve the above problem. TheGabor filter bank was used to convolve with the image and obtained texture featureswhich can be combined with Gaussian weighting function. This method can improve the segmentation performance of the random walk algorithm, can also be a moresatisfactory segmentation results in a more complex image.(3) According to several popular seeded image segmentation methods, researchand extension of the seeded image segmentation framework was put forward in thispaper and also made analysis and comparison of different methods of the frameworkwith different parameter settings. The key parts of the design were discussed in theend.