Research Of Image And Video Segmentation Algorithms Based On Graph Cuts Theory

Image and video segmentation is an important research direction in the fields ofimage processing and computer vision. Meanwhile, it’s also the important foundation ofmany high-level applications, which are based on the analysis of image and video, suchas target tracking and detection, behavior recognition, image and video editing, and soon. All of these need the low-level data characteristics supported by image and videosegmentation. Therefore, the in-depth study of the issue is of great significance andpractical value. Using graph cuts theory to solve the segmentation of the image andvideo is attracted widespread attention of domestic and foreign scholars for the past fewyears.This thesis is based on graph cuts theory, and focused on the binary segmentationof image and video. The concrete work is as follow:1. Summarizing the theoretical approach based on graph cuts for imagesegmentation. The work focus on network flow theory, energy minimization theory, s-tnetwork, maximum flow-minimum cut theorem, the framework of the graph cutsalgorithm and its implementation details.2. Two improvements are proposed for image segmentation algorithm based ongraph cuts. The traditional image segmentation algorithm based on graph cuts uses thegrayscale histogram or brightness histogram to calculate probability for each pixel. Theimproved algorithm uses the Gaussian mixture model to do this, which can processcolor image directly, without conversion; On the other hand, the traditional algorithmuses all pixels to build the s-t network, however, the improved algorithm merges thesame pixels as a node to build the s-t network for speeding up the algorithm.3. Two image segmentation algorithms fusing boundary and regional informationare proposed. GrabCut algorithm can effectively eliminate the background in the object,but it’s poor when the colors of the foreground and background are close; GCBACalgorithm can effectively get the boundary of the object, but it can’t eliminate thebackground in the object. So, an interactive image segmentation algorithm fusingCrabCut and GCBAC is designed in the thesis. The algorithm extracts the boundary of the object by GCBAC to overcome the drawback of GrabCut, and it eliminates thebackground in the object by GrabCut to overcome the shortcomings of GCBAC. Basedon the same idea, an automatic image segmentation algorithm is proposed for thepictures including goods information in a real project. The algorithm gets the boundaryof the object by Canny, and then it can use GrabCut to eliminate the background of theobject. As a result, it provides accurate data source for the next step.4. An interactive video segmentation algorithm based on extended GrabCut isproposed. The implementation details of GrabCut show that the algorithm is an iterativeprocess, it iterates the three steps: labeling, estimating GMM’s parameters and cuttings-t network. The analysis for the performance of GrabCut shows that it’s still consumingtime though the GMM has been stable. Based on the above analysis, the interactivevideo segmentation algorithm based on extended GrabCut is proposed. Throughprompting some pixels in the foreground or background, the GMM can be moreaccurate, and then the segmentation will be good. Through updating the GMM’sparameters between frames, it will reflect the changes in the foreground and backgroundmodel. As a result of the experiment, the respected goals have been achieved.