Improved Level Set Methods For Image Segmentation

Being a basic image technology, image segmentation plays a very important role in image engineering, pattern recognition and computer vision. It separates interesting objects from the background of images, realizes the transform from low-level image data to high-level knowledge, and finally makes the advanced analysis, understanding and artificial intelligence being possible.In the past half of century, image segmentation kept being a persistent hot field concerned by lots of researchers. Although, thousands of algorithms based on different theories are proposed up to now, there still has not a unified theoretical framework to guide the design and implementation of algorithms. Based on the curve evolution theory, active contour model combines low-level image data with high-level knowledge by building an energy functional, and then segmentation result can be obtained as the extreme values of this functional. This way is most likely to evolve into an unified framework for image segmentation. After twenty years of development, active contour model has been well established and became an important branch of the technologies for image segmentation. Under the framework of geometric active contour model, i.e., the active contour implemented by level set, this paper takes still images as main research object and attempts to improve the current level set methods from several aspects, e.g., the robustness against initial curve, a more comprehensive description for images, more adaptive evolution speed and stopping criteria, and more effectively incorporating shape priors constraint. The main achievements of this paper are summarized as follows.(1) A relay level set method for automatic image segmentation is presented. It segments image more times in a series of nested sub-regions that are automatically created by shrinking the stable curves in their previous sub-regions. When the area enclosed by the evolving curve equals zero, a fully segmentation is obtained. The proposed method could detect more inner boundaries and the segmentation result is independent of initial curve.(2) A unified tensor level set methods for image segmentation is proposed. It introduces a 3-order tensor to comprehensively depict features of pixels by combining gray values and Gabor features. This tensor representation provides more comprehensive image features and preserves the spatial position information of pixels. By defining the distance between tensor representations, a tensor level set method is proposed as a generalization of the representative level set method. The proposed method is more robust against noise, improves the segmentation performance for inhomogenous objects, and also can be used to segment texture images due to the introduction of Gabor features.(3) A nonlinear adaptive level set method for image segmentation is implemented. It combines regional information to automatically determine the evolution direction of curve which makes segmentation result independent of the initial position of curve. The adaptive convergence speed makes the curve evolve with lower speed at boundaries and higher speed in uniform region to avoid the boundary leakage. A probability weighed stopping criteria focuses on the boundaries around objects and depresses the ones far away from objects, which contributes to detecting the real boundaries in images. The proposed method combines the edge and region information to reduce the dependency of the initial position of evolving curve, and it could obtain better performance on the objects with ambiguous boundaries and the objects sharing similar gray value with background.(4) A level set method with shape priors for image segmentation is proposed. It utilizes moment-based alignment to deprive of the position, scale and orientation information. That is a non-parameters alignment and does not need iteration to calculate the parameters. Locality preserving projections, i.e., LPP, is employed to project shape priors from observation space to low dimensional subspace in which shape-driven energy functional is designed. To some extent, LPP resolves the problem that the shape priors sparsely distribute in observation space. The proposed method essentially incorporates a constraint, i.e., shape priors, into energy functional, which contributes to a better segmentation performance on the overlapped objects and the objects with complex background.In brief, this paper focuses on reducing the dependency of initial curve, enhancing the representation of image features, avoiding boundaries leakage, and effectively incorporating shape priors. Four associated new level set methods for image segmentation are proposed to improve the current methods. In generally, these research results enrich the theories and applications of level set methods for image segmentation.