Extraction Method Based On Graph Cuts Theory Of Goal

Image segmentation is to separate an image into certain un-overlapped regions. It's one of essential problems in image processing, and a key step from image processing to image analysis. The effect of the segmenting result will directly influent following target expression, feature extraction, parameter measurement etc.The graph cuts theory is an intelligent optimization theory based on the image model of Markov random field and optimal technology of max-flow/min-cut algorithm. It is characterized of ability of obtaining the global optimal solution and unified framework of combining diversified prior knowledge. As a novel instance of energy minimum ideology, it becomes one of the up-to-date research hot spot.Supported by graph cuts theory, this paper focuses on the object extraction in image segmentation field, and following four innovative work has been contributed to such a research:(1) Two object extraction algorithms are proposed based on fast GMM (Gaussian Mixture Model) parameters estimation via graph cuts. GrabCut algorithm enjoys a convenient interactive interface and segmenting ability with high accuracy. However, high cost with estimating GMM parameters is a bottleneck to restrict efficiency of the algorithm severely. The essential of the algorithm is, first of all, to implement iterative estimation of GMM parameters, in order to extract object, so the author put forward the improved version of GrabCut by realizing fast GMM parameters estimation with few representative samples instead of massive pixel samples. Two strategies are presented:①Transform the image into block-image by using the watershed algorithm, and then estimate the GMM parameters with block-image.②By introducing multiscale analysis method, proper multiscale data structure is constructed. The image is pyramidally decomposed into coarse-to-fine scale image series; then GMM parameters are estimated with these image series instead of fixed scale original image series. The algorithm bottleneck is eliminated by the above strategies and results of experiments show that the efficiency of our two methods is 5～6 times as high as that of GrabCut.(2)The author put forward an object extraction algorithm based on IVCN (Inward Variable Contour Neighborhood) and graph cuts, aiming at the limitations of repeated cut of the neighborhood of boundary-reached parts, of overlap between adjacent CNs (Contour Neighborhoods), of failure to extract concave object, and of high possibility of falling into local extreme for the algorithm, etc. GCBAC algorithm is improved within three steps by means of change of CN from width-fixed to width-flexible to eliminate repeated cut of the neighborhood of boundary-reached parts; change of CN dilation direction from bidirectional way to inward way to avoid the overlap of adjacent CN; change of edge weight with weight adjust coefficient to extract concave object and to enhance the ability of noise-overleap of the active contour to decrease the possibility of falling into local extreme. The improved algorithm is much stronger in terms of robustness, much faster in terms of convergence speed and more extensive in terms of its application range.(3)BandCut-a novel object extraction algorithm is proposed by employing image edge information, in which an annular banded region enclosing object boundary with drastic change of the adjacent pixel color value is obtained in an interactive way to realize extraction of color object; the s-t network weights are compensated by introducing the concept of distance gap to define relaxed banded region. BandCut enjoys distinct advantages of no iteration, less computational cost, no local extreme problem, robust parameters setting and extracting ability with rapid speed and high accuracy.(4)On the basis of flat-style s-t network segmenting framework, the author proposed an interactive object extraction correcting algorithm which is characterized by convenient operation, real-time responding, and no damaging to right-extracted parts. Besides automatic correction of local errors, the algorithm also provides a total manual "hard-constrain" style correction way, which makes the algorithm more integrated and efficient.