Research On Spiking Neural Networks And Its Application On Image Processing

As the third generation of neural networks, Spiking Neural Network is the latest research in the field of Neuroscience and Computational Intelligence. It is a dynamical system which can effectively simulate biological neuron of continuous delivery of information over time. The model uses the temporal encoding to organize information. It can simulate biological information processing mechanism, which is closer to the actual biological nervous system than traditional neural network. Studies have revealed that Spiking neurons can deal with external input information by nonlinear processing, whose coding mechanism, neuron model, synaptic learning rules are all very different from traditional neural networks. Studying its internal mechanism and application is of great significance and worthy to explore. As a major branch of the deep learning field, convolutional neural networks can effectively simulate biological brain space hierarchy to interpret the data processing. Therefore, this paper explores that, Spiking biological mechanism can highly imitate pulse sequence processing, while convolution network can extract abstract spatial feature, which are combined with a new way to explore the spatial and temporal model more in line with the new rules of biological knowledge and improve the spatiotemporal data processing capabilities. The subject is not only a strong scientific significance, but also has a high value in engineering.The main work of this paper is threefold: First, we propose a Spiking Convolutional Neural Network(SCNN) model with good spatial characteristics and transmission of temporal information. It takes advantage of Spiking neuron’s timing processing capacity instead of the traditional way to organize, represent and transmit information, for the time pulse convolution. The model combines the Spiking time processing features and convolution cyber space connections. It is based on convolution structure inherits advantages of the local connection, sharing weight structure, highly invariant to data, having less training parameters and so on, along with the computing power of time series data. Learning the best of both, this is the first time that Spiking neurons and convolutional neural networks combine to achieve automatic extraction capability of the spatial and temporal characteristics; Second, this paper presents an image edge detection algorithm based on Spiking-Convolution. It combines two classic, effective filters: Laplacian of the Gaussian(LOG) and Difference of Gaussians(DOG), to improve the traditional convolution methods of image edge extraction. In order to demonstrate the new edge detection algorithm of Spiking-Convolution, we did a simulation experiment, and have achieved good results, providing technical support for further work. Finally, this paper presents a computational model based on Spiking-Convolution mechanism for image recognition. The spatiotemporal information representation model can be divided into Spiking-Convolution scheme, Spiking network structure and learning algorithm. Combination of the three parts offers an efficient and unified Spiking-Convolution identification system. In order to study the performance of the model on image recognition applications, the paper through computer simulation and modeling shows the performance of the algorithm and obtains the desired results. This method enables a high degree of imitation biological properties, in line with the rules of biological cognitive behavior, and can effectively conduct spatial and temporal characteristics of the data when processing identification tasks.