A Level Set Framework Based On Target Shape And Relation Constraint For Image Segmentation

The direction of human social development is to obtain material,energy and information from the environment more systematically and efficiently.With the widespread popularity of the Internet,the explosive growth of visual images quickly broke the cultural environment with text as the main symbol,and became the mainstream information carrier.Image segmentation,as a basic technology in image engineering,has always been the research focus for computer science workers.Driven by energy functional optimization and partial differential equations,combined with the geometric active contour model of curve evolution theory,i.e.level set method,it has a unique advantage in the field of image segmentation due to its flexible and precise target edge presentation ability and the scalability of easy combination of other theories.In real images,there are often problems such as target occlusions,insufficient brightness,and complex background and so on.It is difficult to obtain the ideal segmentation result by relying on the evolution of the underlying image information drive curve,and adding a shape priori constraint to the energy functional is an effective solution.However,the most of traditional level set methods whether with single shape prior or multiple shape priors,basically are for the case where the region of interest is a single target.They could not obtain the desired result when there are multiple targets with similar appearance but different region properties in the images.On the other hand,for the images with similar grayscale between the part of target and the background,or images with deformable target,level set methods would miss the part of target and cause incorrect segmentation results even if the shape priors have been introduced.In this thesis,to tackle the aforementioned two problems,we propose corresponding improvements for level set methods with reference to other characteristics of image target.The main research contents are as follows:(1)A level set method with dynamic shape priors is proposed to handle the concerned issues,i.e.,blur edges and cell adhesion,of cell segmentation in microscope images.Firstly,we place a seed in each cell by preprocessing the original cell image and its LBP feature map,and these seeds are taken into initialization of level set function.Then,fractional derivatives are utilized to establish a new cell-edge indicator for driving the closed curve to cell edges accurately.Finally,we incorporate the automatic initialization and the stopper function into a level set method,and combine with the dynamic shape priors,which is designed by the morphological characteristics of cell,to evolve the curve in common and finally realize cell image segmentation.(2)A level set method based on target shape and relation constraint is proposed,for solving the image segmentation with inhomogeneous and non-rigid deformable target.This work is motivated by the basic idea of pictorial structure models which represent a human target by a collection of parts and connections between pairs of parts.Considering there is no human body parts database,we first propose an algorithm for shape decomposition based on distance visibility to build the required database of shape component priors.Next,for fully expressing the complex human body structure,we represent a part by a level set function on the basis of a multiphase level set framework,and then construct unary shape constraint term and pairwise relation constraint term to drive multiple curves evolving into a semantic shape of human target.The experiments show that our method could improve the robustness of level set methods for segmentation of non-uniform grayscale and deformable target.In summary,this thesis studies and develops the existing level set methods from two aspects,i.e.,the shape constraint of the target and the relationship constraint between shapes.The proposed methods not only provide new ideas with the application of level set methods for image segmentation,but also promote a useful exploration for extending the level set methods to image parsing and image understanding.