Research On Automatic Color Image Segmentation Algorithm Based On GrsbCut

Image segmentation is an important technology of modern information society. With the rapid development of multimedia technology, the demand for color and multi-dimensional image processing techniques start growing. Color image segmentation technology began to occupy an increasingly important position in the professional field. GrabCut algorithm is a novel interactive image segmentation algorithm combines image processing techniques commonly using the Gaussian mixture model and graph cuts theory, which could give a good segmentation. This also makes it a hot topic in recent years.This paper analyzes the GrabCut algorithm graph cuts method for image segmentation theory. The algorithm uses Gaussian mixture model to fit the image of the foreground and background, are classify pixels by tags, and represents the best image segmentation in the form of energy function, using maximum flow/minimum cut theory in graph to solve the function iteratively, to get the best segmentation results. However, the algorithm is initialized manually making the algorithm can not be applied to a wide range.Based on the study of algorithmic processes GrabCut initialization operation, and clustering algorithms by analyzing their similarities and commonalities that when dealing with foreground and the background image segmentation with single object, being initialized manually is similar to 2-class clustering method and thus proposes the use of EM algorithm based GrabCut automatic initialization method. This paper gives a derivation of training Gaussian mixture model parameters by EM algorithm classification methods, and analyzes the convergence of the EM algorithm. Meanwhile, according to EM algorithm sensitivity to initial conditions,this paper proposes k-means algorithm Gaussian mixture model initialization parameters instead of random initialization parameters. To some extent, proposed method avoid the problem of falling into local minima and slow convergence of EM algorithm caused by the poor initialization.Finally, we use modified EM algorithm to replace the original manual initialization, and test results showed that: GrabCut automatically initialized improved in speed has certain advantages, but results from the split of view, improved GrabCut compared with the original algorithm GrabCut accuracy decreased slightly, but the difference was not statistically significant. Basically we achieved an automatic GrabCut segmentation when dealing with the foreground and background image segmentation with a single object.