Research On Constraint Level Set Based Image Segmentation Algorithm

With the high-speed advance and spread of computer software and hardware, digital image processing technology has been developed rapidly, and has been widely used in many fields e.g., remote sensing image analysis, communication engineering, national defense and military areas. As a critical step from image processing to image understanding, image segmentation has become a popular research field in recent years. Segmentation technique divides the image into several parts with different features and extracts the part of interest. The level set based segmentation method transforms the curve in 2D plane into the zero level set of level set function in 3D space, and constructs an energy functional of level set function. The surface is evolved via minimizing this energy functional, and hereby the result of curve evolution is obtained. Because the level set based segmentation method allows the topological structure of the curve to change flexibly, it shows the excellent performance in the field of image segmentation. As a hot research field of image segmentation, the level set method attracts more and more attention of researchers.According to the employed stopping criterion, the level set based image segmentation methods could be divided into three categories, i.e., the edge based level set method, region based level set method and constraints based level set method. Using both the image information and shape priors, constraints based level set method could better handle the segmentation with intensity inhomogeneity, complex background, and occluded objects. However, involving shape priors into energy functional leads an increasing of computation cost, and the efficiency of image segmentation is significantly reduced. In the thesis, three methods are proposed to reduce computational cost and improve the efficiency of image segmentation. The detailed innovative points are briefed as following.(1) A shape constraint based energy functional solved by lattice Boltzmann method is proposed for image segmentation. Firstly, the algorithm uses the method based on image moments to align shape priors. Secondly, the sparse shape priors in high dimensional space are projected to a low dimensional subspace via the locality preserving projections, and a shape-driven energy term is designed in the low dimensional space by a statistical method, i.e., kernel density estimation. Finally, the lattice Boltzmann evolution equation is deduced from the energy functional that is built by combining the shape-driven term with the data-driven term. The minimum value of the energy function is obtained by evolving of the LBM equation, and then the image segmentation is accomplished. Compared with the traditional methods of solving the energy functional, this method can reduce the time consumption and improve the efficiency of image segmentation.(2) A shape priors based salient object segmentation method is proposed. Firstly, the graph based visual saliency method is used to obtain the saliency map of the image, and it is utilized to initialize the evolving curve. Secondly, we incorporate the saliency map into the data-driven term of the energy functional to guide the evolution of curve. The construction of shape-driven term is similar to the first method we proposed in this paper. Finally, we integrate the data-drive term and the shape-drive term to create the energy functional whose local extreme solution is obtained by a number of iterations. Using saliency map to replace random initialization could greatly reduce the time consumption of solving the energy function and improve the efficiency and accuracy of image segmentation.(3) Combining a superpixel algorithm and graph cuts based level set to realize image segmentation method. Firstly, the simple linear iterative clustering algorithm is utilized to generate superpixels, and we select some pixels in each block to construct the energy functional. Secondly, we discretize the energy functional, and construct the corresponding graph by taking a pixel as a node of the graph. Using the min-cut/max-flow algorithm, the minimum cut of the graph is obtained, i.e. the minimum value of the energy function. Finally, according to the properties of the selected pixels, we can get the properties of the superpixels and finish image segmentation. The proposed method is robust against the position of the initial curve and can obtain better segmentation efficiency. It can achieve the global optimal of image segmentation, and can be applied to the batch segmentation of the images with large size.