Method Of Image Segmentation Based On Graph Cuts Theory

Image segmentation is a technology and processes, which is to separate an image into certain regions with special characteristics and to extract the interest foreground. It's the key issue in image processing, pattern recognition and computer vision. And the segmentation result directly affects the accuracy of following image analysis.Image segmentation based on graph cuts is a new approach that developed in recent years. The core idea of the graph cuts is to construct an energy function, then minimize the function by combinatorial optimization techniques. The novelty of the theory lies in its global optimality and unity of combining multiple knowledge.Supported by the graph cuts theory, this paper focuses on the binary image segmentation. Major innovative works are as follows:(1) This paper summarized image segmentation methods in detail, especially the methods based on graph cuts.(2) This paper studied and summarized the graph cuts theory. Besides, it described the basic knowledge involving in graph cuts. Including s-t Network, Network Flow theory, s-t Cut, Maximum Flow/Minimum Cut theorem and so on.(3) By researching and analyzing GrabCut algorithm, this paper described the Gaussian Mixture Model, which is used to characterize the probability distribution of color information. And described its theoretical basis of mathematical and its wide range of applications.(4) Supported by GrabCut algorithm, a fast image segmentation method based on graph cuts is proposed in this paper. GrabCut algorithm based on graph cuts has the global optimality and the unity of combining multiple knowledge.However, the algorithm is less efficiency because of using the whole pixels to initialize the GMM parameters and using iterative algorithm to obtain exactitude. The essential of the GrabCut is:first, iterative learning, training GMM parameters; then extracting object from the original image under the identified GMM. The purpose of previous iterative cuts is to determine the GMM, which prepare for extracting the target later. The last cut is the real object extraction. So we learnt that there is no need to use all the pixels of the image as samples, select a few representative samples also can achieve the goal. So, we proposed an idea that using less samples to replace the whole pixel samples. The strategies is:transforming the image into low-frequency images and high-frequency images by the wavelet transform, taking low-frequency image's pixels as the samples of GMM estimation, applying graph cuts used in GrabCut algorithm to determine GMM parameters iteratively. Finally, cutting the original image according to the established GMM and the target image was obtained. The experiments showed that our method significantly improved the algorithm's efficiency in the premise of accuracy segmentation.(5) GCBAC (Graph Cuts Based Active Contours) algorithm uses boundary information to extract object, it can't penetrate into the internal of the object boundary. So it is powerless for images that objects have holes. GrabCut algorithm uses the GMM to express the probability distribution of background and foreground color information, it uses regional information to extract object. It is a good segmentation algorithm for the image whose color information varies greatly between foreground and background, but when the foreground is very similar to the background, the extracted object is not accurate, it needs a large number of amendments, but artificial amended image is often not accurate, so it often brings error to the further analysis and processing. As a result, for some images that objects have holes or object's color information is similar to background's, objects can't be extracted precisely only by boundary information or region information. For these defects, combining boundary information and regional information simultaneously, an object extraction algorithm based on graph cuts was proposed. The experimental results showed the algorithm can extract such objects precisely and effectively for images with holes in its objects or images with similar foreground and background. The method achieved the expectation with high accuracy in automatic segmentation, less user's workload and high efficiency.