A New Segmentation Model Of Computer Vision Based On Biological Vision

Foundational theories of computer vision (CV) have brought CV great progress, but the more progress it gains, the much more problems it is showing. Despite of the frames of Marr's Vision computational theory or later attentive vision theory, they couldn't resolve segmentation problem, which is still ill-posed problem, and their segmentation hypothesis models are poor and improper in segmentation. Foundational theories of CV have almost exhausted themselves in CV application. Thus this thesis just explores original CV theory for successive CV development. We have in detail examined all the theoretical foundation in CV from AI, suggesting that it hardly dodges the limitation like to AI paradigm limitation in research: general knowledge formalization in traditional AI and segregation of figure-ground in CV are the same problem, which essentially is the problem of global & local recognition, and traditional formal method can't formalize the whole. Segmentation or formalization is only the method of reconstructing the whole using local information, a kind of perceptual reductionist approach.We claim that research on biological vision (BV) will offer methodology to CV research and application. So we've extensively and deeply investigated segmentation from point of views of BV, a large view, and CV itself, relative less view, putting forth two different levels of models: BV Holistic Model and Global-to-Local Feedback Perceptual Model, GLF Model.The Holistic Model illuminates that 1) Existing and competition is the purpose of BV being and evolving, and develop with interaction between itself and environment. 2) Vision is a dynamical learning process. 3) Visually perceptual forming is from-global-to-local multi-feedback result. 4) Visual attention plays important roles in visually spatial search, segregation of figure and background, knowledge feedback from high cortical level andvisual consciousness forming. The mechanism of visual information processing is different completely from that of computer processing. And the former learns environmental knowledge around by implicit paradigm, not by formalized logical deduction, which deals with information in a from-local-to-global way, and adopt from-global-to-local multi-feedback approach inverse to formal methodology. This is the problem which the trouble of traditional AI and CV lies in.Following above, we've investigated segmentation and things reflecting it with a obvious view of the problem. We remark that all various of segmentation algorithms are based on a certain hypothesis about segmentation, such as similarity, proximity hypothesis etc. segmentation hypothesis is a kind of local description of general character of outer object. Presently, CV just reconstructs the holistic description of special object from the local description. Hube and Wiesel established hierarchical theory of visual receptive field is the right premise of segmentation hypothesis, which resolved the existence and constructable problem of perceptually neural mechanism. And where construction is an inverse segmentation under recognition. We think constructablity of the premise and formalization go with each other intrinsically: the both are from-local-to-global reductionist approach and can't escape the difficulty mentioned above. With this problem, we suggested a GLFM according to third rule of the holistic model. The GLF model shows us that perception is not just the single processing from parts to whole, but comes into genuine, holistic perception about object through multiple whole-parts-whole feedback processes, among which attention is critical to modulate early and later information processing.We simulated the process of perception forming under the condition of attention through experiments, which shows proof that final stable perceptual forming is the result of multiple interation between outer perception and inner transcendental knowledge.