Research Of Computating Model Based On The Biological Vision Characters

There are 80% stimuli which are processed by visual system among the human brain sensory information. Visual information processing mechanisms have become an intense study in psychology, neuroscience and computer science for recent two decades. To research the information processing methods in the optic nerve system is studying how the human brain responses and deals with the stimulation from the outside world. The key point of understanding the way of brain working in is to study how the brain represents,encodes,processes the stimuli outside and how to explain the world.Based on the physical characters of the received field of the human visual neural cells, a model is proposed in this thesis to probe the way in which the human perceive the outside world. Besides, on the basis of efficient coding hypothesis and manifold learning theory, we improve the LLE to model the way the human recognize the world.Firstly, a contour detector with directional surrounding suppression, motivated by biological principles, is proposed in this thesis. The orientation preference of the visual neural cells and the interaction effects of adjacent cells are analyzed in this model. The Gabor filter is used to simulate the simple cell. The Gabor energy filter and edge orientation map are briefly explained. Then,■directional surrounding inhibition based on DOG filter and directional DOG filter is proposed.Secondly, based on the efficient coding hypothesis and manifold learning theory,an improved LLE------LLE with adaptive neighbors is proposed and applied in objectreorganization. We first analyze the disadvantage of LLE, and then against the fixed neighbors, the selection criterion for the adaptive neighborhood is proposed. We also use the RBF kernel distance to replace the Euclidean distance, and give a simple and effective method of mapping the new samplesThe proposed model with directional surround suppression is applied to detect contour. The experimental result shows that contour detection by using this model is more effective than by using the classical contour detectors and this model can distinguish better the different texture boundary, isolated edge and object contour. We also apply the improved LLE to recognize objects. The result shows that, compared with the normal LLE, the improved algorithm is perform better in object recognition such as manifold expanding, data classification and pattern recognition. This proves the effectiveness of the algorithm.