A Neural Networks Based On Competition In The Application Of Invariant Object Recognition In The Visual System

The research on cognitive process of human visual system has great significance in machine vision and cognitive science, especially now that the two famous theories in this field, which called respectively the theory of early characteristics analysis and the theory of early global perception, are still totally different. In this thesis research was conducted by focusing on visual system to study corresponding visual phenomena with the method of the neural network model based on biological mechanisms. The main work and some highlights of this thesis were summarized as follows:1. When a stimulus is presented, preliminary processing takes place in retina where basic features are extracted and an electrical representation of the image is transmitted as neural spikes along the optic nerve. The signal travels to the Lateral Geniculate Nucleus via two pathways, parvoellular and magnocellular pathway. The LGN acts as a relay switch in the brain and routes the signal to the back of the brain towards the visual cortex. The object recognition takes place in the ventral pathway .2. The previous modles use the concept of alternating simple cells and complex cells in the visual cortex to build invariance about the target object. The model implemented for this thesis is a neural network where the trace learning rule modifies the weights between connections to build tolerance to changes in size, location, or rotation of the object.3. The model will be tested for invariance in rotation, translation and scale. Then it will be tested to see if the model is able to respond using different categories. The original image and various images with rotation, transition or scale, which have been trained by the neural network model, after comparing the trained results, verify the model can simulate the visual invariance properties well. Also explain some of the visual phenomenon which have no authority conclusive, finally we find to the modle can make a good performance when we apply it to the object category.In a word, research on visual system was a complex issue. The neural network model proposed in this thesis theoretically not only conform itself to relative characteristics of visual system in theory, but also explain some relative visual phenomena in practice, especially made a good effect on describing visual cognition and simulating human visual characteristics, which would have the possibility to unify the theory of early characteristics analysis and the theory of early global perception in further research.