Research On Object Recognition Methods Based On Scene Relative Information

Object recognition is the hot spot in the field of computer vision research. Directed toward the shortcoming of traditional methods in the field of object recognition, the paper researches a method that can model huge amount of objects in uniform forms which is called the object recognition method based on scene relative information, and analyse the recognition model and the parameters in this method. It explains the emphasis and the nodus of this method is the gaining of shape primitives. The paper deeply researches the method of gaining shape primitives. The following is the main work:1. Applied 3 widely used methods in image recognition fields to gain shape primitives. They are the method based on features block, line detection based on endpoint connection, line detection based on supportive regions respectively. The method based on feature blocks can extract blocks with complete edge. The limitation of it is that it demands abundant feature blocks with homogeneous distributed characteristics of objects. The method of line detection based on endpoint connection extract lines through the micro-mechanism of lines and has a trait of quick and simple. The method of region supporting don't need edge detection and has a strong ability of noise-immune. In the article, it explains a method of vanishing point line groups detection based on line detection. Used to recognize those objects that related to building structures and have a large amount of parallels, it is simple and quick.2. Analysed the method of gaining shape primitives based on polygon approximation by concrete instance. Use the camber characteristic of curves to constraint the species group space of GA in traditional method which is based on global characteristics. It accelerates the velocity of convergence of the arithmetic. Definited a constraint condition of the connective sequence of vertexes of approximative polygons directed to the problem of the unset shape of them caused by diverse connective sequence. Definited a form of approximate error and raised the a method of circumference-maximize and reduce the runtime of the arithmetic. Explained the theory of controlling the approximate scale by adjust the error threshold. Different approximate scales meet the request of different resolution.3. Raises a method of window-based vector quantization for primitive division and a method based on RBF neural network for primitive classification directed to gain arc primitives from natural images and analysed the affect of 3 main parameters of it. The method of window-based vector quantization don't need complex operation and for this reason it is quick and simple. Experiments registered that it can reach an ideal result when the parameters are in proper values. It provides an feasible method for gainning arc primitives.