Research On Auto-Determination Of The Parameter Values Of Pulse-Coupled Neural Network

Artificial neural networks plays the key role in the fields of artificial intelligence and image processing because of it's imitation of the wisdom and power of biological nervous system. As a new kind of neural network, Pulse-Coupled Neural Network (PCNN) has been put more emphasis on the research in many fields since it was proposed by Johnson in 1994. PCNN model was created by the research and understanding in the visual cortex of small mammal such as cat and produces synchronous bursts of pulses from neurons with similar activity, effectively grouping them by phase and pulse frequency, which makes it significantly different from the conventional artificial neurons. Just because the model derives directly from the studies of the mammal's visual properties, PCNN has been used successfully in image processing fields.PCNN model is a algorithm with multiple parameters, and finding the proper value of these parameters is an onerous task. The drawback of the PCNN hinders its further development and application. So many simplified PCNN model is adopted in many papers. The remaining parameters are, nevertheless, manually adjusted according to the concrete application. The high accuracy is unachievable with the simplified model and the quality of latter study is affected.Up to now, no papers which investigate the auto-determination of the parameter values of PCNN model in full-range and systematically exist. Based on the mentioned above, two algorithms which can performs the auto-determination are presented in this research. As one part of this paper, genetic algorithm which is a general purpose stochastic optimization for search problem is utilized to determine these parameter values automatically, and an automatic PCNN system is established. Through the segmentation, the performance of the automatic PCNN system is verified. As another part of this paper, gradient descent algorithm is adopted to search parameters which can reduce the error between the desired output and the actual output gradually according to the LMS principle, then a self-adaptive PCNN system is established. Given only an input and the desired output, the adaptive PCNN system can find all parameter values necessary to approximate the desired output.The correctness and dependability of two algorithms are verified by experiment results, the good foundation for the latter study of PCNN model in image processing is established.