Representation And Application Of Object Prototype Based On Visual Perception Mechanism

Object representation and recognition is an important topic to imitate capa-bilities of human’s visual perception. But the task is too difficult, and to achieve better results, the implementations of computer applications need to draw lessons from ideas of human’s visual perception mechanism. Human’s visual perception is a system created by the evolution during billions of years. Although researchers of artificial intelligence and computer vision have made great achievements on object recognition in recent years, their artificial systems are still much less effective than human’s visual perception mechanism. So, we need to hunt for inspirations from the theories of neurophysiology and cognitive psychology on visual perception.This thesis designs a prototype representation which conforms human’s recognition mechanism and has much intelligibility, and a training algorithm framework by combining the viewpoints of some pattern recognition theories on cognitive psychology, such as recognition-by-component theory and the topologi-cal theory of vision. Firstly, the designed algorithm uses mixture of 2-dimensional Gaussian component to fit the object’s edge images, drawing lessons from the viewpoint of recognition-by-component theory. Secondly, the thesis gives a sub algorithm to match components between two samples, and based on it, design a cluster algorithm to mining components and components’structures which fre-quently exists to fuse this series of feature information into one prototype repre-sentation. This prototype representation is different from the others in traditional bottom-up recognition methods on computer vision and draws lessons from the topological theory of vision, since the prototype uses the structure of graph or tree to describe the object’s large-scale structural and topological information, which makes the representation obtain perception features of top-down recognition con-forming characters of human’s recognition while being used on recognition.To verify the intelligibility of the prototype and whether it conforms charac-ters of human’s perception, this thesis works for rebuilding the component image reversely from the fused information when the prototype has been built, to give a visualized expression of training result. Moreover, it explores the application on recognition by the prototype. It uses the template image rebuilt from the proto-type to find the coarse-grained candidate positions of the object, and takes the component-level recognition which is advanced and has the character of top-down processing into consideration then. In particular, it gives a formalize framework description of component-level recognition to convert the question to a optimiza-tion problem based on an energy function, and preliminarily designs some search strategy to detect the object’s candidate positions.