Research On Closed Contour Extraction Guided By Visual Perception

The task of computer vision is to construct computational models of human vision and to simulate the process of human visual perception. Perceptual organization, which organizes the low-level features to be meaningful perceptual objects, is a significant part of human visual perception system. As a manifestation of perceptual organization, the objective of closed contour extraction is to obtain the contour of perceptual objects and identify the salient structures in images. It is one of the important ways for defining and obtaining perceptual objects. Closed contour extraction is the fundament of middle or higher visual tasks such as object detection, object recognition, shape matching, etc. It is also very significant for information processing like deepening the understanding of perceptual objects, semantic analysis and memory storage. Designing reasonable and effective algorithms of perceptual object contour extraction based on the understanding of cognitive mechanism is one of the core issues of computer vision.This dissertation is mainly focused on designing closed contour extraction algorithms to simulate the vision information processing. According to the three-level information processing of computer vision, all the studies are based on the research results about visual perception in cognitive psychology and neurophysiology, and take Statistic and differential geometry as tools for analyzing the image data. The studies of this dissertation can be divided into three parts. The first part aims at improving the input quality of contour extraction. In this part, we design and implement the boundary-point detection algorithm based on intensity and texture features. This part realizes the simulation of low-level information processing of vision. The second part explores two ways of closed contour extraction. One is an edge detection and linking method and the other one is a deformable model method based on energy function. This part realizes the simulation of middle-level information processing of vision. The third part describes the topological structure of closed contour by simulating the short-range connection among cells in brain cortex as a network, and implements the shape matching. This part realizes the simulation of high-level information processing of vision. The main innovation points of the dissertation are as follows:First, we propose a boundary-point detection algorithm based on maximum difference of features in natural images with complex texture and low-contrast boundaries. This algorithm uses local binary pattern as boundary-point detector, chooses optimal threshold by minimizing the combination of image conditional entropy and the reciprocal of simplification complexity. According to the maximum criterion of gray difference in multi-scale, the algorithm computes contrast value of each texton and cuts down the insignificant edge points. Compared with other algorithms, our algorithm can reduce a large number of background and texture information which is useless for contour detection, at the same time save salient boundary points. As a result, this algorithm provides a much clear and simple boundary-point image.Second, aiming at the problem that edge detection results are sensitive to noise, we propose a salient structures fitting algorithm based on contrast value of boundary points, then use the salient structures as fitting edges join in the contour grouping process for closed contour extraction of perceptual objects. This algorithm uses segments to fit density regions in salient boundary-point images. The optimization goal is a binary function based on distance between points and segments as well as contrast values of points. In the split step, a adjust split process is proposed whose goal is to evaluate the quality of the fitting. The projections' density of neighboring points is computed in this process, which makes the segment models have more proper length. The algorithm uses fitting edges as input, combines the closure as cues, extracts closed contours by choosing the global structure dominant contour grouping model. Experimental results show that binary optimization goal can locate the segments more swiftly, fit salient structures more accurately. The fitting edges improve the efficiency of contour grouping process by providing more refine and simple input.Third, the contour grouping methods for closed contour extraction have religious foundation in physiology and psychology. However, the modeling of Gestalt rules is hard to be universal, and the substance of grouping is only to cluster the edges and get discrete segments of the contour. Aiming at this, we use continuous models to replace the discrete segment models in contour grouping methods, and propose a curvature-dependent magnitude active contour model. Based on salient boundary-point images, this model fits the object contour by the curve evolving guided by global salient information. According to the global salient templates which provide cues for the moving direction, the model is robust to initialization. The magnitude of external force is a function of curve curvature which makes the model flexible as well as self-interaction avoidable. The redefined energy function, with salient boundary-point images as input, enables the model to manage both intensity and local texture information in images. Experimental results show this model is robust and efficient in failure cases of other models, and able to gain good performance on fitting the closed contours of perceptual objects in low-contrast natural images.Forth, aiming at the problem that local shape descriptors are sensitive to noise, occlusion and variety of illumination, we propose a global structure shape descriptor based on responsive contrast, and build the similarity based on the combination of topological structure and local information for shape matching. The global structure of shape is described by the statistical law of differences in position and orientation of segments, which is inspired by the short-range connection among cells in neurobiology. In the shape matching step, this algorithm preserves the order of shape features by considering not only the global structure but also the punishment about the local information differences. Experimental results show the descriptor is robust to translation, scaling, rotation and occlusion in a certain extent. The results accord with the attributes of short-range connection and the responsive invariance. Even in extreme occlusion circumstances, the shape matching has similar results with those of human vision perception.