Modeling And Extracting Color Feature Through Visual Pathway Using Spiking Neural Network

The human visual system can process a variety of color vision information and show excellent performance.Understanding and simulating the biological color vision are an important research content in computer vision field,and the developed biologically inspired computational models have been widely used in image processing and pattern recognition.Therefore,on the basis of understanding the neural mechanism related to color information processing and its formation,we build a hierarchical spiking neural network model to simulate the color information processing from the retina,the lateral geniculate nucleus to the primary visual cortex and secondary visual cortex through plasticity learning,and implement the color feature coding.The main work and innovation in this thesis are as follows:(1)The spiking neural network and its stability are studied.Spiking neural network is often called the third generation of artificial neural network,with better biological plausibility and higher information processing capacity.First,several typical spiking neuron models,spiking neural network and synaptic plasticity are introduced.Then the stability of existing spiking neural network and its spike-timing dependent plasticity(STDP)learning are analyzed and their problems are pointed out.Finally,in order to solve these problems,we emphatically study and simulate the impact of inhibitory synaptic plasticity and several homeostasis mechanisms on the stability of spiking neural network.(2)According to the functions,structures and response modes revealed by the neuroscience,including multichannel characteristics of color vision,single-opponent receptive fields,double-opponent receptive fields and neuronal lateral connections,the four-layer model of the Spiking Neural Network is built to simulate color information processing along retinal-LGN-V1-V2 color visual pathway.In this hierarchical model,the connection properties at each stages of the pathway are achieved,and in addition to STDP learning rules,two neural mechanisms including homeostasis and inhibitory synaptic plasticity are added to make the neural activity of color vision processing more stable.(3)The performance on the hierarchical network is evaluated.First,including the hue map in V2 cortex obtained by our hierarchical network resembles that of physiological experiments.Second,the experimental results on encoding features of color-biased images by using the V1 or V2 channel output in the hierarchical network show that the V1 or V2 channels simulated in the hierarchical model have different degree of color constancy.Finally,the classification experiments on the SFU Lab datasets using the V1 or V2 feature coding further verify the validity of the color features extracted by the hierarchical network.