Based On The Classification And Identification Of The Mechanism Of The Human Brain Visual Perception

The human vision system can easily identify different kinds of objects whereas it becomes quite difficult for computers. As a hot and difficult problem in the field of computer vision, object categorization and recognition has been received extensive interests with broad prospects. However, the lack of systematic guidance currently on the type of image features and how to extract features from image has left computer facing a big challenge, when dealing with tasks such as object recognition in clutter backgrounds and categorization in different backgrounds. Therefore, to build a feature computational model as human visual information processing mechanism referring to the research results from cognitive psychology and neuroscience is becoming an attractive research direction.Based on human visual cortex, the model for visual feature computing is studied. Furthermore, we attempt to improve the standard model by Serre, Wolf and Poggio in several different ways. The main work in this thesis can be summarized as follows:Firstly, some basic physiology researches and the selective attention mechanism are introduced as well as the method of salient point, Gabor filters and Support Vector Machine.Furthermore, the standard model by Serre, Wolf and Poggio, and the sparse feature model with limited receptive fields proposed by Jim Mutch are introduced and analyzed. In order to make more effectiveness in using training image and simplify features so as to decrease the complexity of computing, a series of improving methods are proposed and introduced. The new sparse feature model is integrated with visual selective attention mechanism.Finally, the model proposed in my thesis is tested on Caltech 101 object categories and AGG database for object categorization and object recognition in clutter background. The new model is analytically compared with the standard model proposed by Serre and Jim in both correct rate and the speed of training as well as feature dimension and training image's effect.The results show that our improved model can effectively extract features from images and has good learning capability even on relatively few training images and in low feature dimension space.