Rearch On Spiking Neural Networks And Its Application In Image Segmentation

Spiking neural networks (SNN) are the most realistic model, which often bereferred to as the third generation of neural networks. Compared with the firstgeneration of neural network disposing information with threshold-gates and thesecond generation of neural network expressing information with activation function(e.g., Sigmoid function), the Spiking neurons communicate with each other usingprecise time of spike emission, taking into account the biological phenomenon like LTP,LTD and STDP. This makes the information processing, neuron model, learning rule ofSNN quite different from the previous two generation of ANN. It’s very meaningful tostudy the internal mechanism and applications of SNN.Image segmentation is one of the most basic problems in computer vision field.The focus of this article is applying SNN to image segmentation. This article firststudied common models of spiking neurons, pointing out the intrinsic dynamicproperties of them through simulation of the model, which is very useful forinformation processing.Learning rule is a critical issue to influence the performance of SNN. Accordingto the characteristics of image, this thesis proposes a learning rule for weightadjustment of spiking neuron using to encode the pixels’ gray value of the image,which can handle the spatial-temporal information. The weight adjustment rulecombines STDP with winner-takes-all, taking into account the neuron neighborhoodinformation at the same time.Network architecture is very powerful for applications of SNN. This thesisdesigned a three layer feedback networks for image segmentation. The image codinglayer using population neurons to encode image gray value, which output the fire timesof excitatory and inhibitory neurons to feature binding layer. There are lateralconnections of neurons in feature binding layer, taking into account the neighborhoodof image. The image segmentation layer has the same size neurons to the number ofthe needed segmentation regions, which have a feedback connection to feature bindinglayer. Eventually based on synchronous time binding characteristics of Spiking neurons, we will rearrangement the image matrix, and get the final result of imagesegmentation.According to the above network architecture, this thesis presents a design methodof model object of SNN under the MATLAB simulation environment, including thedesign of structure and function properties of the network object. Taking into accountthe importance of the winner neuron, this thesis store the fire time sequence of neuronsin a priority queue. For simulation the entire network, this thesis proposes aclock-driven algorithm for the simulation of SRM neuron and an event-drivenalgorithm for the simulation of weight adjustments between two neurons.Finally, this thesis gives a simulation of image based on the proposedsegmentation method. By comparing the segmentation result with traditional and theLEGION method, we confirmed the effectiveness of the proposed algorithm. Issueslike parameter optimization and network convergence need to deepen in furtherresearch.