Study On Visual Perception Inspired Image Segmentation System

Image segmentation is a general concept,which includes special image segmentation in image processing fields and object segmentation in computer vision fields.The former is aim to separate the image into several regions by their specific features,the latter means to figure out the semantic object from other uninterested objects and background.As people are often interested in the specific regions in image.Image segmentation,especially the object segmentation,is always the one of the most fundamental and difficult techniques in image processing and computer vision.As an important image analysis technique,image segmentation provide quantitative information for object recognition,image measurements and image understanding, which has a very broad application in static and dynamic images processing.The traditional segmentation pays more attention on the regions with similar features or specific features.Distracted by the complicated image scene and imaging noises,it is still very difficult to arbitrarily separate the object from its background.The most obvious obstacle is derive from the great variant in the inner parts of object,the object region can not be extracted efficiently and will not bring about information for successive stages like object recognition and measurements.Like other image processing techniques,image segmentation is improved as an interdisciplinary approach addresses engineering and other research areas.Among them,the psychological and physiological researches on the visual perception system have close relation with image segmentation.This thesis gives the concept of segmentation a new understanding from the view of visual perception inspiration.Learned from the procedures of biological visual activities,this thesis give analysis of the categories and the development of segmentation techniques.Some novel algorithms are completed in this thesis related to both the Bottom-up and Top-down segmentation methods,which have close relation to visual perceptual inspirations.In this thesis,some aspects related to the visual perception inspiration have been involved;there are multi-scale perception,visual attention and u(?)ified mechanism of segmentation and recognition using Top-down & Bottom-up procedures.The corresponding computer technology are the following fields: construction of nonlinear scale space and segmentation application,visual attention based and region competition controlled image segmentation,hierarchical Top-down & Bottom-up object segmentation system.The level set method is also researched as a Bottom-up approach.The new tendency of modern image segmentation is to develop Top-down, Class-specific segmentation techniques.The main characteristic in the new approach is to introduce segmentation process into the object recognition process,complement to Bottom-up segmentation.Correspondently,these Top-down segmentation methods have been supported by biological functions from both the feed forward and feedback visual pathways existed in the biological visual systems.Object prior and object model is necessary,and the object model,simulating the cognitive and perceptive activities of prediction and decision-making,drives the segmentation.To overcome the segmentation difficulties mentioned above,it gets rid of the restriction of typical clustering techniques and is developed to based on underlying element blocks from the class object units,this method is characterized by a simultaneously segmentation and recognition procedures.One of main work in this thesis is to set up a hierarchical tree to represent the object model,tree nodes are those patches or sub-patches as local features model.The information of shape,appearance and position are included in the tree nodes.During the training stages,the criteria of extraction of patch are related to gray and appearance features,the object skeleton is also obtained as the topological structures. Based on the skeletons,patches are matched to the image and obtained the object shape and positions hypotheses.The Top-down similarity map is derived from the process for the purpose of object estimation.Beside the Top-down results,the cooperation from the Bottom-up methods are needed to give the region similarity map, all those maps are used for object hypotheses maps to help to segment the object.Faced with the complicated visual world with different level of feature densities and complexity,the biological visual system can handle those situations by multi-scale and multi-resolution sampling,so that the object can be described accurately and the perception capacity is improved robustly.Researchers in the field of image processing also notice this,as a result,the multi-scale image analysis has been formed as a mainstream technique and can be applied to almost all image processing fields.Scale space is divided into linear and nonlinear scale space,the implementation of scale space is linear and nonlinear filtering.Different degree of filtering gives rise to different level of scales.Using the isotropic nonlinear filtering can smooth image while preserving edge and object main structure,the nonlinear scale space are constructed by this way to analyze the detailed and global structure of object stage by stage.Another work in the thesis is to speed up the construction process of nonlinear scale space in order to provide structure information for multi-scale image segmentation.Contrasted with the previous methods which using total variation model in the scale space,this thesis pay more attention on the research of scale parameters selection and suggest the scale total variation concept and to modify the previous algorithms.The scale selection is extended from temporal domain to spatial domain.The scale parameter is now changed to filtering neighbor with different sizes. The disadvantage of previous method is derived from the diffusion iteration number, as parameter is time-consuming.The new parameter changed back to spatial parameter can bring some advantages like the filtering process time reduction and being more consistent with the general scale concept.This thesis introduces the basic theory of filtering used in the scale space and analyzes the selection of parameters,so the option of parameters can be extended to more aspects.The stop condition is also analyzed.The application of nonlinear scale space has demonstrated in the image segmentation and image filtering. The seed points are very important in region based image segmentation techniques.For example,the fuzzy connectedness based segmentation is a practical method applied in medical images.But the difficulty is to set the multiple seed points in the image to guide the segmentation.The general methods take advantage of clustering analysis,but these approaches are often inefficient as the influence of image fuzzy features.If there is no explicit localization,it has to consider both the object and its background regions to help the judgment of the pixel whether it is belonged to object or background.Actually,this method is still belonged to explicitly giving seed points location.From the analyses of visual performance,visual attention on object is valuable for the recognition and separation of object from others, although there are many criteria used to quantitatively guide the selection of seed points,the analysis from the view of attention mechanism is seldom.The other work in this thesis is to propose a new method of seed points localization.Guided by the visual attention model,the seed points are equal to the visual salient points.Through the research,the localization of saliency points is connected with the estimation of image feature probability density function,the new visual saliency map is introduced by this thesis to consist the computational mode of visual attention.By the non-threshold method,the object salient features are extracted to produce seed points to guide the segmentation with region growth.The efficiency of new model is greatly improved than previous model through quantitative description,one of the application areas is the fuzzy medical image.It can localize the seed points in fuzzy image automatically and segment image with accuracy.Multiple seed points can be controlled by the region competition and speed up the region merging process in the fuzzy connectedness based segmentation.Recently years,the level set function is applied to the active contour model in image segmentation and become a classic approach in image segmentation.Active contour model is rooted with the original Top-down segmentation concepts,the current level set method is also developed to combine more high-level statistical priors with lower level features.Active contour model is continued to be developed by combining Bottom-up and Top-down procedures.The more influence from the Top-down,the more robust in the object segmentation,so there are still more potentials in the level set methods.In this thesis,level set is researched from the Bottom-up view,the Chan-Vese model is modified and this thesis proposes new method based on the multiple regions information.As for the weak edge and leakage problems,the modified Chan-Vese model can enhance the contrast between regions,to reflect difference and to help the region competition.The weak edge problem is solved by the classification of different object regions,which uses more global features,rather than simple growth or shrinkage of regions,which is only based on local features.The image is decomposed into sub-regions successively and sorted into a similar binary tree structure.The advantage of similar binary tree used in region representation is to avoid the merging process with no splitting process at later stage in bigger regions,which is not considered in general region competition.During the merging process,a balance between different types of features is established as the energy criteria,the gray value fitting is cooperated with more emphasis on curve smoothing.