Research Of Image Segmentation Methods Based On Deformable Models

As a primary task of image engineering,image segmentation has been widely used in medical science,military,public security,transportation,etc.Image segmentation has always been the focus of computer vision,and thousands of segmentation methods based on different theorems have been proposed.The deformable model based on curve evolution treats image segmentation as a calculus problem,i.e.,minimizing the corresponding energy functional.Besides,it provides a unified framework for image segmentation,and facilitates the introduction of high-level prior knowledge.Therefore,the deformable model has been deeply studied and widely applied.According to the form that the curves are represented,the existing deformable models could be categorized into explicit models and implicit models.The later,i.e.,the level set method(LSM),encodes the 2D evolving curve as the zero level set of a level set function defined in 3D space,by which the topological changes of curves could be dealt with in an elegant way.Besides,the associated energy functional with respect to a level set function is optimized to realize the segmentation results.However,the deformable models only involving image information could not obtain desired results when the images with objects occluded or complex background come upon.According to the number of involved shape priors,level set methods with prior knowledge could be grouped into two subgroups,i.e.,the one with single prior and the one with multiple shape priors.The key idea is to design associated regularized term and incorporate it into the energy functional to constrain the evolution of curve.Generally,single prior could not reflect the commonality of shapes,so the level set methods with single prior could not achieve desired results,when the object differs from the prior even they belong to the same kind.Meanwhile,for the case of multiple shape priors,it is not always guaranteed that sufficient shape priors are available in any scenarios.In such a way,the vectorized shape priors sparsely distributing in a high dimensional observation space results in the difficulty to obtain an accurate estimation of distribution,and reduces the effectiveness of the regularized term.In this thesis,we propose two methods to tackle the above problems as follows.(1)A level set method with multiple priors based on independent component analysis is proposed.First,we project vectorized shape priors from the observation space into the low dimensional subspace in which kernel density estimation is utilized to estimate the probability distribution of shape priors.Then,the regularized term based on the shape distribution in low dimensional space could constrain curve evolution,and realizes the segmentation of the object with specified shape.(2)A hybrid level set method with semantic shape constraint is proposed.Inspired by related techniques of shape matching,we propose a novel hybrid level set framework by incorporating shape context as a semantic shape descriptor.The designed shape constraint term based on shape context makes sure that the proposed method flexibly model the shape deformation with same semanticity.Although,only involving single shape prior,the proposed method could achieve competitive and even better performance in fewer number of iterations.Additionally,the hybrid framework contributes to incorporating other shape descriptors,and facilitates the extension of level set methods.In summary,the two proposed methods improve the performance of conventional level set methods with shape priors.The second method,in particular,innovatively provides a new hybrid framework to incorporate semantic shape constraint.The two proposed methods not only enrich the theory of level set methods,but also contribute to the application of level set methods.