Research On Object Segmentation Method Of Binocular Images

The aim of image object segmentation is to extract the object information from the given image.After the segmentation is completed,each pixel in the image is assigned a label which is the foreground(object)or background.With the gradual maturity of stereo camera technology,the increase in the data volume of binocular images makes the target segmentation of binocular images become one of the research directions in the field of computer vision.In a general sense,the object and the background are generally at different depths in the image of the natural scene.Compared with monocular images,binocular images contain the depth information of the images,so using the image depth information can not only better describe the objects in the scene,but also more effectively express the foreground and background information.How to use the depth information to enhance the segmentation effect is the key to the object segmentation of the binocular image.So the difficulty that needs to be resolved is how to represent the depth information in a mathematical model,and how to add the resulting mathematical representation to the object extraction process.Based on the interactive segmentation of graph cut,this paper improves the effect of segmentation by adding the depth information of the image.First,the user marks a rectangle that contains the object in any image from the given binocular image.The pixels inside and outside the rectangle are used to generate the foreground and background prior models respectively.Then,a graph model is constructed by the foreground/background prior model,the similarity of adjacent pixels points in the image,and the correspondence between the pixels in the image.Finally,the object region is extracted by iterative graph cut optimization.Compared to the traditional method of defining the foreground/background prior model and neighborhood similarity in RGB space,the method of this paper adds the distribution of foreground/background disparity information to enrich the prior model and defines similarity measure of adjacent pixels in RGB-D space.The depth information is used to two ways in our method.On the one hand,according to the depth information which has certain clustering characteristics,the depth information is added to the graph model in the form of a probability model and is used to certain weight value together with the RGB probability model to calculate energy function regional term.On the other hand,a similarity calculation method of image pixels in an RGB-D space is defined,and the obtained similarity is added to the graph model for the calculation of the boundary term in the energy function.The work in this paper shows that the effect of image object segmentation has been improved by adding the mathematical model of depth information since the advantage of depth information in binocular images over monocular images.This work has important reference significance and potential application value for further study of computer binocular vision.