Segmentation Algorithms By Random Walk

Image and video segmentation is an important research topic of computer vision,the ambition of image and video segmentation is to segment the region of interest in the image or the video for subsequent processing analysis.Image segmentation and video segmentation algorithms have been widely used in monitoring system face recognition and so on.This paper focuses on efficiency,accuracy of the image and video segmentation.The main contents include the superpixel segmentation based on adaptive nonlocal random walk and video segmentation based on spatio-temporal random walk.Superpixel is a very important preprocessing method of computer vision,which is an indispensable part in many image and video processing method.For example,it will be difficult if using the pixel as the basic unit to compute the relationship between two frames,but if using the superpixel as the basic unit,it can greatly reduce the time and space complexity.The existing superpixel segmentation algorithms can not segment the weak boundary well,but random walk theory is a very powerful model which has a good performance on dealing with complex texture and weak boundary,it can improve the performance of the segmentation of the weak boundary.Some of the existing video segmentation algorithms don't have a good spatio-temporal consistency,the segmentation accuracy of such algorithms is insufficient and the speed of the algorithms is not enough.Applying random walk method in the video segmentation can get the segmentation results with good spatio-temporal consistency and accuracy,and it cost less time to segment the video.The main work of this paper is as follows:(1)Applying random walk method into superpixel segmentation,this paper proposes an superpixel segmentation algorithm based on adaptive nonlocal random walk(ANRW).The traditional superpixel segmentation doesn't consider about the relationship of the adjacent pixels,our adaptive nonlocal random walk algorithm make some improvement on the traditional nonlocal random walk algorithm(NRW),and this paper also proposes a new method to choose seed points based on the gradient.The experiments are done on Berkeley dataset(BSD300),it shows the effectiveness of our algorithm.This paper also extend the adaptive nonloocal random walk algorithm to interactive image segmentation.(2)We propose an semi-supervised video segmentation algorithm using spatio-temporal random walk.Our algorithm need the user need to mark the object we want on the first frame,and then using spatio-temporal random walk to segment the next frame,and we finally optimize the initial results by improved sub Markov random walk,our algorithm has accurate segmentation results and is efficient.The comparative experiments on SegTrack and the additional experiments on SegTrack v2 demonstrate the performance of the algorithm.