Researches On Image Understanding Based On Hierarchical Visual Perception Mechanisms

Image understanding is the current hotspot and difficulty in computer area, which is to interpret the image scene and its contents by using computer systems and realize the behavior of understanding the outside world as the vision systems of human and other higher organisms. Image understanding and computer vision both have the aim of studying and embodying the visual cognitive ability of human, thus the researches on image understanding from the perspective of human visual perception system has important theoretical significance and application prospect. It has been proved that the distinct level property of image understanding is highly consistent with the hierarchical perception mechanisms of human visual system, therefore, the understanding of structure and function of visual perception system and the construction of corresponding mathematical models are important means and fundamental starting point, which can extend and develop the existing image understanding methods.Considering the relationship between image understanding and computer vison, image understanding and cognitive science, this thesis focuses on the representation, learning and understanding of visual information based on the research findings of visual physiology and psychology. We mainly analyze the hierarchical visual perception mechanisms and their corresponding computational models. Combined with the dominant mechanisms of structure and function in visual cortex, a novel image understanding framework based on hierarchical visual perception mechanisms is built, which can implement a series of visual task such as image segmentation, object contour detection, generic object recognition and scene classification in different levels.The main work of this thesis includes:(1) The development history, research status and level property of image understanding are summarized. The physiology structure, functional characteristics, and related research results are also introduced. Furtherly, we discuss the computational models based on hierarchical visual perception mechanisms, and point out several related structural and functional mechanisms in visual cortex. Finally, the relationship between image understanding and human visual perception system is analyzed and then the research framework of image understanding based on hierarchical visual perception mechanisms is proposed.(2) The description of low-level feature in image understanding is studied. Aiming at the key issues of color image segmentation, we first construct a group of hierarchical visual feature that describes the visual consistent image regions, and then the categories of pixels are classified by utilizing the interaction between bottom-up and top-down module inside the Fuzzy-ART model. Combined with the specific strategy of image region merging, a hierarchial visual segmentation model is constructed. (3) Based on the topological connection in visual cortex, we furtherly research low-level visual description in image understanding. Through analyzing the lateral inhibition and topological connection between neurons and their vision formation mechanism, topological structure and self-organizing learning strategies are introduced into the original ART model, and thereby a topology preserving ART model is constructed. Under the constraint of topology preserving, the new ART model can describe low-level visual properties and perform object contour detection and region segmentation better.(4) According to the representation of middle-level structural information in image understanding, we analyze the existing models of object recognition in visual cortex. By virtue of the abstract information model in primary visual cortex and the mechanism of sparse coding, a kind of sparse representation of object ptototype can be formed, based on which we present a hierarchical feature extraction method for generic object recognition.(5) Taking the high-level visual cognition as the background, we analyze the fast formation of perceptual Gist representation of scenes based on the effect of global and quick perception of scenes. Dynamic characteristics of visual cortex and its relation to holistic scene perception are researched, and we propose a computational model for scene classification based on Gist features derived from global perception in human visual cortex.