Pulse-Coupled Neural Networks Based Vision Inference Model And Its Application

Since recent years, the high-speed railway has been rapidly developing in China. With the highly increase of the speed of trains, the guarantee of operation safety is becoming more important. For some random factors, among the factors affecting the safety, like foreign object invasion, and line damaged caused by disasters or human, there is no effective monitoring and prevention methods, because of their strong randomness."Vision based surveillance is a hopeful method for these factors, but the present surveillance systems or vision based detection systems for railway line usually can’t detect automatically. This is mainly caused by the constraints of current computer vision technologies. Although computer vision has made great progress in recent years, it still far always from its ultimate goal, which is to make the computer have same ability as the human visual system. Recently, using clues from biological vision research to solve computer vision system has become an important direction. And the Bayesian inference on Markov random field (MRF) has been widely applied in both computer vision and biological vision research. In computer vision Bayesian inference is a uniform paradigm for solving many practical vision problems, but it’s usually difficult to solve. Existing methods cost a lot of time, and can’t be applied to real-time processing systems. In biological vision, in other hand, Bayesin inference can explain many psychophysical findings, which means the brain may using similar mechanism for vision information processing. Furthermore, the brain can process vision information very efficiently, which suggests that neural network can quickly solve Bayesian inference benefit from its parallel architecture. However, the neural implement mechanism is still unclear. Based on these issues and take visual inspection of high-speed railway line as the application background, this thesis proposed two algorithms for two core issues in railway line surveillance:foreground detection and stereo matching. Both of the two algorithms are based on Baysian inference paradigm. And this thesis also proposed a biologically plausible neural implementation for Baysian inference on Markov random fields, which provides a useful mean to solve Baysian inference problems and also a plausible guess to reveal how biological vision implement Baysian inference.Firstly, this thesis proposed a biologically plausible neural network implement of Bayesian inference, which is called neural inference model. It can perform Bayesian inference computation on general Markov random fields. The model is a pulse-coupled neural networks, whose basic elements are leaky integrate and fire (LIF) neurons, the information passed between neurons is represented by pulses. The neural inference model perform Bayesian inference based on belief propagation algorithm. The architecture of the network is introduced, the network is composed by two kind of neurons:belief neurons and message neurons, which compute believes and messages respectively. The belief and message are encoded by population code, so the analogy value is represented by average spiking rate of a population of LIF neurons. Simulation result shows that the proposed model can perform Bayesian inference efficiently.Secondly, based on the neural inference model, for two core issues in railway line surveillance:background modeling and stereo matching, two algorithms based on Bayesian paradigm are proposed. For background modeling, a new likelihood using neighborhood information and a background update method combine long term and short term update mechanism were suggested. The algorithm is tested by wildly used dataset, and is compared with other algorithms, the performance of the proposed algorithm is better then recent published algorithms. For stereo matching, a MRF based matching algorithm is proposed. The likelihood of the MRF model is computed using outputs of intersecting cortical model (ICM). The algorithm are evaluated using testbench wildly used, the result show that the proposed algorithm give better result then recent neural network based method.Finally, the vision detection system for high-speed railway line is constructed. The system have two subsystems:foreign objects inspection system and fence completeness system. The architectures, methods and detection result is introduced in the thesis.