Based On The Shape Of A Priori And Image Segmentation Graph Cuts

Image segmentation is separated from the object of interest in an image, which is making the image subdivided into a number of meaningful sub-regional or object. It is one of the basic problems in image processing and computer vision. Image segmentation results in a direct impact on the latter part of the image analysis and processingImage segmentation methods based on graph cuts can achieve a global optimal solution in binary labeling problem,and can get a local optimal solution with strong features in multiple-labels problem.The basic idea of image segmentation based on graph cuts is to construct an energy function based on the image information,then the image segmentation problem is transformed into energy function minimization problem to solve. The specific approach is to construct the network figure with the point of the image pixels as the the vertices of the network according to the image’s own information, the figure makes energy function regarded as the a collection of of the edges in the graph, so that the value of of the energy function is corresponding to the cut capacity of the the network figure, and then the use the maximum flow/minimal cut algorithm to get the the minimum value of the energy function, so as to get the optimal solution.Graph cuts is generally only to take into account of the image smooth and edge information.For the images which have simple background or have greater differences between the target and the background,you can get better segmentation of the target;but if the images have clutter background,noise or obstructions, due to the large interference of external factors, the segmentation of target is incomplete or contains background part. To overcome this limitation, you can consider adding some other constraint information, such as a prior of the shape, making the algorithm contain more constraints, you can get better segmentation of the target image.In this paper, we research in the image segmentation problem based on shape priors and graph cuts, the mainly work done in the following aspects:(1)Firstly,image segmentation method based on a single shape priors and graph cuts is studied, adding shape prior knowledge on the basis of the graph cuts algorithm, and let the algorithm contains more constraint information, thus the algorithm can restrict the object search space by more constraint and extract the target completely. Segmentation method based on image information has poor results,whenthe images have noise, background clutter or target block. So adding prior information of the shape, and it should add prior shape which is exactly the same to goals you want to obtain, and describe the prior using the shape distance discrete description, then incorporate the prior information into the figure through the terminal edge weights, and finally use graph cuts to minimize the energy function. The results of experiment show that the algorithm has the robustness and higher accuracy.(2)If between the given shape prior template and the splited target,it has translation, rotation, scaling or scale affine transformation, in order to take advantage of the template of the shape, we use Sift (Scale invariant feature transform) feature matching algorithm for processing, according to the correctly matching feature points, then use the affine transformation formula to obtain the transformation parameter and thereby given template is aligned with the target of the segmentation, to solve the affine differences between the affine shape prior template and the segmented target. Experimental results show that algorithm after considering the affine transformation is more flexible,and able to respond to the situation that has affine change between shape prior template and the target to be segmented.(3)On the above studies, the use of the template is a single fixed target shape template that can only deal with specific images, having limitations for other similar images. So we study the image segmentation method based on kernel principal component analysis (Kernel Principle Component Analysis) nonlinear shape priors and graph cuts. The method uses KPCA to train a group of shape template to obtain a shape statistical model, resulting in the shape of a prior, and then combining with graph cuts for image segmentation. The experiment showed that the prior based on KPCA has flexibility, and the flexibility different from the type mentioned in (2) is no longer confined to a fixed template, but to use a plurality of shape templates to prcocess the different images, having a better result in spliting similar goal of similar shape.