Research On Binocular Image Object Segmentation Method Based On Graph Cuts

Image segmentation is a process of classifying the pixels in an image.After years of development,the traditional monocular image segmentation technology has been very mature which can accurately segment foreground objects in the image.At present,binocular stereo vision technology is gradually maturing,the number of binocular stereo cameras is increasing,and the amount of binocular image data is growing rapidly,which makes scholars pay more attention to the binocular image segmentation.Binocular image segmentation is becoming more and more important in the field of computer vision.Generally,the foreground and background of an image are in different depths.Therefore,the foreground and background of an image can be better distinguished through the depth feature.Compared with the traditional monocular image,the binocular image contains the depth information of the captured scene.Therefore,how to effectively exploit the depth information in the binocular image to improve the segmentation effect has become the key to the binocular image segmentation method.This thesis introduces depth information into the Graph Cuts image segmentation method to improve the image segmentation effect.The main research contents are as follows:(1)A neighborhood enhanced graph cuts method for binocular image segmentation is proposed.Firstly,the color feature and depth feature of the binocular image is integrated,and the foreground and background probability models of the graph model are constructed by the integrated feature.Secondly,three neighborhood systems of the graph model are constructed.(a)Find out adjacent pixel pairs in the image and construct a traditional neighborhood system.(b)Find out correct matching pixel pairs through disparity image left-right consistency check,then,we construct the cross-view neighborhood system.(c)For each pixel,find out first k pixels that are most like it in the feature space,then,we construct the neighborhood system of the feature space.Finally,a graph model is constructed,and the energy is minimized by the Graph Cuts algorithm to obtain the final segmentation result.Experimental results show that the segmentation results of this method are more accurate.(2)A multi-feature selection binocular image segmentation method based on Graph Cuts is proposed.Because the effect of the method in(1)is not good at the edge of the foreground object,a multi-feature selection segmentation method is proposed to solve this problem,which transforms the image segmentation problem into a multi-classification problem.First,the foreground and the background probability of pixels in the binocular image is calculated based on the color feature,texture feature and depth feature,respectively.Then,we construct the foreground\background probability model.Secondly,according to the characteristics of the three neighborhood systems of the method in(1),we modify them respectively to adapt to the multi-class Graph Cuts framework.Then,we construct the neighborhood system of the graph model.Finally,the graph model is constructed,and the energy is minimized by the multi-class Graph Cuts algorithm to get segmentation results.The experimental results show that compared with the method in(1),this method has a better segmentation effect at the object edge,and the overall segmentation result is better.