Research On Computational Model Of Contour Grouping Guided By Global Saliency Information

Perceptual organization is an important mechanism of human visual perception system. It attracts many researchers who employ themselves in neurophysiology, cognitive psychology, computer science and other fields in recent years. The process of perceptual organization is a link between the optical signals in natural world and the perceptual objects. It is also the basis of many high level visual tasks, such as image recognition, attention assignment, memory storage and so on. Contour grouping, which takes image edges as grouping elements, is an important part of perceptual organization. The contour grouping model can find salient structures in images, and is an important tool for the definition and obtainment of perceptual objects. It can be seemed as the foundation of constructing object-based attention model, object detection model and object recognition model. Researching the cogni-tive and neural mechanism of perceptual organization to design a contour grouping model with high efficiency which accords with human perception system is indis-pensable in the study of visual perception system inspired information processing theory.Applying the research results about perceptual organization in cognitive psy-chology and neurophysiology into computational models is the key of constructing contour grouping models which accord with human perception characteristics. In this dissertation, we propose the quantized model of Gestalt grouping cues and con-struct a contour grouping model guided by global saliency information. All of these studies are based on the Gestalt grouping rules and the research results in psychol-ogy and neurophysiology, take Statistic as the tool, and mainly emphasize the role of attention in perceptual organization.According to the three stage of constructing a contour grouping model, the study of this dissertation can be divided into three parts. The first part aims at improving the input quality of contour grouping. In this part, we design and realize the natural image edge detection algorithm which is suitable for contour grouping. The second part analyses the statistical properties of the natural images, and give a reasonable definition and quantization of gestalt grouping cues between directed tangents. The third part defines the grouping cost based on the role of attention in perceptual organization, gives the optimization of the grouping cost, and finally realizes the contour grouping in natural images. The main innovative points of the dissertation are as follows:First, we propose a multi-scale boundary detection algorithm based on global saliency information. Compared with other algorithms, our algorithm can get better detection results, which are more proper for contour grouping. Nonlinear combina-tion of the edge detection results in multiple features can ensure the accuracy and integrity of the edges. The spatial prior getting from the boundaries of salient re-gions can remove noises and trivial edges effectively. Linking neighbor edge pixels to form edges with a certain length can make the detection results more steady and proper for contour grouping. We update the spatial prior and trace each edge across multiple scales, and evaluate the saliency of each edge according to the edge energy existing in its whole life.Second, we propose a new similarity measurement for directed tangents in con-tour grouping, which is called directed edge region measurement. We develop a graphic interactive tangent label tool, and by which we construct a standard and accurate human labeled directed tangents set. We discuss the key factors which impact the similarity measurement and give the best parameters according to nor-malized variation standard for similarity measurement on labeled directed tangents set. The directed edge region measurement for tangent similarity measurement over-comes the indefiniteness of tangent groups, which can improve the importance of similarity cue in contour grouping. Comparison with other similarity measurements on the tangents set shows that the directed edge region measurement is much more accurate.Third, we construct a generative grouping cue combination model for contour grouping. This model makes no assumptions of the independence of grouping cues like other discriminative models, which can lead to a more satisfying description of the statistic properties of gestalt grouping cues. The generative model is a more exact gestalt cue combination model than other models. We use the proximity as a key grouping cue, and discuss the joint probability distribution of continuity and similarity in different proximity conditions. Using the generative cue combination model, we fit the joint probability distribution of continuity and similarity on human labeled set which has a special form, and give a more accurate description of the correlation of gestalt cues.Forth, we propose a contour grouping algorithm guided by global saliency infor-mation, and realize the object-based attention model based on the hierarchical con-tour grouping algorithm. We use the global saliency information to guide the process of contour grouping on the basis of the research achievements about the attention role in perceptual organization in psychology, and obtain the contour grouping of salient perceptual objects in natural images. We consider both the same object ef-fect and inhibition of return in attention, and realize the hierarchical attention shift among perceptual objects using hierarchical contour grouping model. The experi-mental results on natural color images show that our contour grouping algorithm can group the salient perceptual objects and is more effective than other grouping algorithms in most images. Also, the attention shifts of object-based attention are hierarchical and are agreed with human visual perception.