The Analysis And Research Of Neural Networks Used In Key Stereo Technologies

As a new field based on various interdisciplinary, the theory research and practical application of the computer vision has made great progress. Especially in recent years, the development of image processing technology and high performance computer coming into being provide favorable conditions for computer vision research and application. So, the further research to computer vision has much theory meaning and practicality value. The stereo technology is a significant branch of computer vision, and shows great potential applications towards practicality. So the stereo technology gets more and more important station in the whole computer vision.The thesis is based on Marr vision theory, expatiate domestic and overseas development and achievement on binocular stereo vision. Based on analysis and conclusion of diversified methods, we analyzed matching and calibration of key stereo technologies systematically.After analysing and comparing with the existing stereo mathing thearies and methods, the thesis adopted the Hopfield network to match. The basic thought of algorithm: the energy function is built on the basis of uniqueness, compatibility and similarity constraints, which reflects the constraint relations of all pixels of the same lines. It is then mapped onto a 2-D neural network for minimization, whose final stable state indicates the possible correspondence of the matching units. Because there are noise and ptical distortion in images, the matching ffects are not ideal used Hopfield network with only three constraints. In the thesis, the constraints of location and depth are added into the energy function, and show the better effects.The thesis introduced which is the theory and the present conditions of domestic and international research of the stereo-camera calibration, and carried on the analysis to the existing techniques of this area and the difficulty to calibration. After studying the knowledge of the traditional calibration, self- calibration and active calibration, the thesis carried on the research of stereo calibration based on neural networks. To avoid the deficiency of standard BP network on training time and generalization, the new training algorithms based on function backwards constricting and Bayesian generalization method is adopted and good results have been acquired respectively. Besides, In the thesis, a neural networks-based and traditional calibration-based method is presented. It is expected that, with this approach, we can maintain the major advantage of linear methods and obtain improved accuracy without...