Research On Shape-prior-based Variational Sparse Segmentation Model

Using the computer to simulate the function of human eyes to accurately segment the objects is very important for both computer vision and image processing.Considering that the objects in real images might be influenced by noise,occlusion or similar background,using only the image-based information to segment these objects often leads to poor results.One effective way to improve the performance is to combine the over-completed dictionaries derived from the prior shapes with the sparse representation models to guide the segmentation.Based on a deep survey and study of many classic sparse shape representation models and varational segmentation methods,this thesis focuses on the research of the sparse shape representation models,the shape dictionary generation methods in the projected space and the varational segmentation models combined with the sparse representation.In addition,several shape-prior-based varational sparse segmentation models have been proposed based on the research above.The major work and contributions of this thesis are as follows:1.Based on several propositions,we deduced that by extending the definition of the projected coefficients,we could build a convex shape set with a sparse subset based on the shape projected function.Using the sparse composition of the elements in the sparse subset to constrain the input can assist in further building the varational sparse segmentation models on the projected shape space.The studies on independent component and non-linear kernel shape space have indicated that the modeling method above had a certain generality,and it also provided a new way to build the varational sparse segmentation models in the projected space.2.Based on combination of the varational level-set methods,the independent shape component and sparse shape representation,we proposed a sparse independent component representation based variational object segmentation method.This method solves the problem that the validity of recovered shape cannot be guaranteed when the independent components are used directly to formulate the sparse shape combination.Moreover,by replacing the original shape by the log-polar shape,the proposed method also solves the automatic shape aligning problem in the sparse shape representation.Through building the shape set based on the space spanned by the independent shape components and iteratively solving the sparse representation of the level-set function,we eventually formulated a sparse independent component representation model to supervise the varational object segmentation.3.Based on the recent proposed "wake-sleep" method,we built a hierarchical segmentation framework.This framework separated the energy optimization into two phases:a "wake" phase and a "sleep" phase,wherein the "wake" phase mainly focused on the representation of the low-level image information,and the "sleep"phase centered on strengthening the high-level representation of input shape.Meanwhile,a dualization term was also formulated to balance the energies from both the "wake" phase and the "sleep" phases.With these techniques,an implicit kernel sparse shape representation based object segmentation model was proposed to solve the problems in searching the sparse neighbors in the non-linear kernel space and the implicit kernel shape representation supervised varational object segmentation.4.In some applications,the sparse shape representation based segmentation may encounter the small sample size problem,and it will limit shape representation ability of sparse shape combination.To solve this problem,we proposed a dictionary-group-based varational sparse segmentation model.The model added a local constraint together with a weight coefficient to achieve a flexible local adjustment for the shape.Compared with the traditional method,the proposed model fully exploited the shape information of the training set and largely improved the representation ability of the shape-prior-based varational sparse segmentation model.It also solved the problem that the varational sparse segmentation model cannot recover the original shape of the object,when the object had local deformations.5.To improve the shape deformation ability,we proposed two shape modeling method:a distance constrained probabilistic shape model and a fuzzy log-polar decomposition shape model.The former added a constraint to change the probability of the original shape over a range of distance.The distance constrained probabilistic shape largely extended the connotation of the training shapes,and it also could simulate the property of the "intermediate belt",which commonly existed in the average shapes.By extracting the local information deep inside the training samples,the fuzzy log-polar decomposition shape model improved the segmentation accuracy and provided more rich information of the shapes,especially for the small sample size problem.We tested our models on both public datasets and self-built datasets.The experimental results showed the superior segmentation capabilities and robustness of the proposed methods.