Research On Interactive Image Segmentation Based On Graph

Image segmentation is a technology and processes of extracting the interesting regions, which is to separate an image into a collection of many disjoint regions with different characteristics of pixels. It is a branch of image processing, and the segmentation results directly affect the research and development of pattern recognition and machine learning and so on.Image segmentation based on graph cuts is an approach that attracts many researchers’ attentions in resent years. It constructs an image model based on Markov random field, then minimize the energy function relied on Max-flow and Mini-cut algorithm, finally obtains the segmentation results through constructing mapping relationship between image and graph.Supported by the graph cuts theory, this paper makes some researches and improvements from different aspects. Major researches and innovative works are as follows:(1) This paper summarizes image segmentation methods in detail, especially the rational theory, developments and improvements based on graph cuts.(2) This paper summarizes radical knowledge of graph cuts, it includes Markov random field, network flow model, Max-flow and Mini-cut algorithm and the mapping relationship between image and graph and so on. By researching and analyzing Grabcut algorithm, this paper mainly describes mathematics theory and extensive applications of Gaussian Mixture Model.(3) In the constructed network graph based on Grabcut, the proportion of every node belongs to the foreground or background is only related to its own characters. All of these will make a weak relevance among pixels, and finally give a wrong mark to a part of regions. Especially the eye region in the face image, it has obvious differences with its surroundings. When we use the Grabcut to segment face image, it is easily determined as background and results in incomplete segmentation results. To improve these problems, an algorithm combining the optimization of weights and CS-LBP texture feature is proposed in this paper. First the multi-scale watershed is applied to pre-segment the original image into regions to construct region adjacency graph. Through this step, it can reduce the quantity of the node in the network graph. Then it extracts color and texture feature from each region, and enriches the feature information in every node. Designing an iterative optimized algorithm to correlate the weight of data item of a region with its surroundings and the texture constraint is added to the energy function with adaptive parameter. Through these improvements, the final improved algorithm is applied to facial image segmentation, and obtains better segmentation results. The experiments contrasted with former show that segmentation efficiency is improved in the algorithm proposed.(4) Traditional shape prior method need people adjusting the constraint extent of the shape prior constantly when the images include noise or shelter. For one thing, it will increase the complexity of interaction. For another, if it is relied on more shape prior model, it is prone to produce segmentation errors. Because every pixel in the image contains different intensities of noise and color feature. Responding to these problems, the adaptive shape prior method is proposed to constrain the process of image segmentation through adjusting the weight of shape prior item. At the same time, the Grabcut algorithm has a high time complexity in the period of Gaussian Mixture Model parameter evaluation, it is also an important factor that restricts Grabcut algorithm developing rapidly. The proposed Gaussian Mixture Model with muti-sampling in this paper evaluates the parameter relied on minor pixels, it can reduce time consumption and improve the total speed of traditional algorithm. The experiments show that the improved algorithm have better robustness, more accurate results, shorter time consumption and extensive application, when dealing with the image contained noise or shelter.