Research Of Developmental Network Model Simulating The Working Mechanism Of Visual Cortex

Object recognition is one of the core issues of computer vision. Compared to computer vision, the human visual system can quickly and accurately identify the target in the field of vision, so in recent years, there exist many object recognition models simulating working mechanism of human visual system. This paper presents the developmental network model, which can complete target recognition in an image,and then simulates two important pathways, the ventral pathway and the dorsal pathway of visual cortex to build a model imitating visual cortex's working mechanism and finishing the target identification under the complex background. In this paper, as an example of object recognition, however, face recognition research is not the key and the ultimate goal is to verify the effectiveness of the network model in object recognition.The basic principles, algorithmic process and the internal competition mechanism of the development network are introduced, and difference from the traditional network is also reported.A basic developmental network model is built to test the ORL face database. We compare and analyze experimental results with two types of initialization of weight, images initialization and random initialization. The best recognition rate is over 95%.Next, in order to detect face data with complex background, synapse maintenance is added to the model to deal with the target areas of images, which can weaken the background. The experimental results show that the improved model is valid and effective and the best recognition rate is over 96%.Finally, Where-What Network based on developmental network is introduced.The characteristics of the model are firstly proposed, and then the internal parameters,such as receptive fields, neurons states and calculation process of the network are described. By experimental design and analysis, weights are visualized to observe inside working case easily. The results of experiment prove that not only type, but also location and size all can be effectively detected. The best recognition rate canreach 100%. Location error can be reduced to 1 pixels and the minimum size error is close to 1 pixels. The validity of the model is verified by the results.