Method Study, Interactive Image Segmentation Based On Graph Cuts

Image segmentation is an essential and key step in image engineering, which result will directly influent following image analysis and image understanding. Taking Markov random field as probabilistic model of image structure description, max-flow algorithm as optimal technology, the graph cuts theory constructs an energy minimum based image segmentation technology.Supported by the graph cuts theory, this paper focuses on the image segmentation field, and follow innovative work has been done to such a research.(1)A fast segmentation algorithm based on watershed transform and graph cuts is proposed. GrabCut enjoys high segmentation ability in a convenient interactive way. However, its segmenting efficiency is severely low because of its iterative image cutting way. We improve GrabCut by the ideology of estimating Gaussian Mixture Model parameters with few typical samples instead of enormous pixel samples. Our method transforms the image into block-image with watershed algorithm, and then estimates the Gaussian Mixture Model parameters with block-image instead of original image. In such way, the algorithm bottleneck is eliminated and the segmentation efficiency is sharply improved.(2)A fast segmentation algorithm based on multiscale analysis and graph cuts is proposed. With the same guide as (1), the image is decomposed into multiscale image serials from coarse scale to fine scale by means of multiscale analysis method, the Gaussian Mixture Model parameters are evaluated with such image serials instead of fix-scale image serials. Our method eliminated the estimation bottleneck and obtained notable segmentation efficiency.(3)A real-time correcting algorithm via graph cuts is proposed. To correct the possible local errors in the preceding segmentation result, an independent post-stage correction method is provided, which has the advantage of handy operation, instantaneous response, exact protection to the extracted parts, and two correcting way of manual and automatic. It's a strong perfecting to the preceding segmenting.