Research On Rapid Interactive Image Segmentation Method Based On Graph Cuts

Image segmentation is an important part of image processing, and it has been widely applied and plays a major role in many fields such as computer vision, image editing, pattern recognition. In this thesis, the interactive image segmentation technology is the main focus. Taking use of the priority information provided by the user, interactive image segmentation methods extracts useful objects that the user is interested in from the original image through related algorithms. The quality of segmentation result and the efficiency of segment method play a decisive role in the later image operation.Graph theory, which is completely different from traditional segmentation methods, provides a new research area for image segmentation technologies, and it has cause great attention by many researchers. Firstly, traditional gray-scale image segmentation methods are introduced in detail in this paper, followed by the concept of graph theory and the similarity between graph and image. In the last, the method that transforms segmentation of image to segmentation of graph is analyzed in theory.In graph theory, graph cuts algorithm is one of the mostly used algorithms in interactive image segmentation technologies. The basic framework of the graph cuts algorithm is analyzed in this paper, as well as a series of related applications. By comparing the advantages and disadvantages of these methods, some improvements are made in this thesis.GrabCut and Lazy Snapping are well performed methods in interactive image segmentation methods based on graph cuts algorithm. The quality of images has been greatly improved with the development of digital technology, traditional interactive segmentation methods show more and more defects in the calculation time and segment quality. With help of pre-segmentation method in Lazy Snapping, simple linear iterative cluster algorithm which incorporates multi-scale structure tensor is used for pre-segmentation of the image, combined with GrabCut framework for later segmentation operation, a rapid image segmentation method based on superpixels is proposed in the thesis. After pre-segmentation of the image, using superpixels to describe the characteristic information of the pre-segmentation area, replacing statistical characteristic method in GrabCut method, then a Gaussian mixture model is established related to the simplified image, expectation maximization algorithm is applied to learn, update the parameters of the Gaussian mixture model build. In the last, Max Flow/Min Cut theory is used to cut the s-t weighted graph which is mapped from the cost function, a rapid segmentation of the image is finally achieved.Through comparison of traditional GrabCut segmentation method and GrabCut segmentation method base on superpixels, the results show that the segmentation efficiency is greatly improved under the same segmentation quality. The improved method is feasible and practical, and it can be widely applied in daily life and professional fields.