Research On Image And Video Segmentation Methods Based On Graph-Cut Theory

Image segmentation is an important part in the field of computer vision.The goal of image segmentation is to classify the pixels of the image in a meaningful way.There are many algorithms of image segmentation.This paper mainly discusses segmentation methods based on GraphCut theory.The Graph-Cut theory maps the whole image into a weighted undirected graph,and the pixel is regarded as a node.The edge between nodes is broken into subgraphs,so that the internal similarity of the subgraph is kept to the maximum and the similarity between subgraphs is kept to the minimum.There are many segmentation algorithms based on Graph-Cut theory.There are many methods of image segmentation based on Graph-Cut theory.This article attempts to propose some new algorithms based on three methods:GrabCut,OneCut,and SnapCut.Our first contribution is to propose a GrabCut image segmentation algorithm based on local sampling.This method no longer clusters the global seed points K-Means and establishes GMM,but conducts a local search near the pixels to be segmented,takes the searched foreground seed points and background seed points as samples,clusters the foreground and background of the samples according to the binary tree structure,calculates the penalty for classifying each pixel to be segmented into a leaf node of the binary tree,and finally completes the segmentation through the maximum flow minimum cut.The experimental results show that this method has certain advantages when the color difference between foreground and background is not obvious.Our Second contribution is to propose an OneCut image segmentation algorithm based on texture features.OneCut proposes an optimized energy function.This method converts OneCut to Lab color space and calculates the L1 distance of color histogram as the entropy function.At the same time,it proposes to use the texture feature information of the image to segment the edge of the image.By introducing the texture feature operator LBP into the boundary item,it compares the texture features of the edge pixels,magnifies the texture difference between the pixel at the edge of the image and the center pixel,and obtains a more obvious segmentation penalty.The experimental results show that this method has some advantages when the color transition between the foreground and background edges of the image is smooth.Our third contribution is to propose a multi-frame bi-directional constrained SnapCut video segmentation algorithm.The local window classifier of SnapCut needs to track the target boundary between multiple consecutive frames.This method proposes to modify the position of the classifier of the next frame based on the motion estimation of this frame by referring to the spatiotemporal information of multiple adjacent frames before and after the super-pixel position information matched by the adjacent frames before and after being mapped to the optical flow field to compensate the original optical flow vector of the pixel,At the same time,the foreground probability of the color model inside the local window classifier is linearly weighted with reference to the foreground probability of the local window classifier in the adjacent frames before and after,so as to avoid the occurrence of jumping points and obtain a smoother probability.The experimental results show that the proposed method has certain advantages when the video object is mixed with a lot of changing background due to its motion.