Research On Nonlinear Retinal Coding Base On Multimodal Stimulation Input

The retina is the first organ for information processing and coding in advanced animal visual circuits and it can convert input light stimuli into biological spike sequence.The study of the physiological processes of retina ganglion cells provides fundamental technical support for the construction of artificial retina coding models,which is prerequisite for artificial vision.The existing works on retinal coding models are mostly based on artificial white noise stimuli,which lack adaptability in natural and complex scenes,and cannot simulate biological characteristics of retina.Therefore,this thesis studies and realizes a nonlinear computation retinal coding model based on the information processing mechanism in retina.The main content of this thesis is as follows:This thesis first studies the physiological characteristics of retina ganglion cells,and proposes a non-linear model based on multi-modal stimulus input.This model firstly extracts the temporal and spatial characteristics of the input video based on the specific response mechanism of the ganglion cells to natural visual stimulus,which simulates the preliminary processing of the upstream cells in retinal circuit.Secondly,this model introduces the historical excitation state of neurons and horizontal stimulation of coupled neurons,which composes the physiological input of retinal ganglion cells.Afterwards,the model uses stimulation basis functions to compile the input stimulus into a multi-modal stimulus input matrix,which has both bio-interpretability and advantages in computation.In the neuron internal computation,a pair-wise filters convert input stimulus into positive and negative inputs to simulate the excitation-inhibition balance regulation mechanism in neurons.Finally,the excitatory-inhibited stimulus inputs are summed and activated by spike nonlinear calculations to obtain the firing rate of neurons.Experiments show that the multimodal stimulus input nonlinear model can accurately simulate the natural video coding in retina ganglion cells,and the performance is remarkable better than other nonlinear retinal encoding models.Secondly,to further study the characteristics of biological retinal coding,this thesis analyzes the computational characteristics and physiological functions of biological ganglion cells in the visual formation process,and proposes the corresponding theoretical analysis and experimental verification on the reliability,variability,regional sensitivity and coupling mechanism of ganglion cells.Experiments indicate that the retinal coding model proposed in this thesis can realize the visual coding function of biological retina and reproduce the physiological function of ganglion cells in retina.It is a nonlinear retinal coding model with both retinal coding and biological features.In summary,this thesis proposes a multi-modal stimulus input nonlinear model for retinal coding for natural scenes.Experiments show that our proposed model has high accuracy,sensitivity and certain biological characteristics,and it shows significant advantages in comparison with the classical non-linear computing coding models.The proposed model in this thesis not only can expand the application scenarios of retinal coding models,but also can help to build more advanced models for retinal coding mechanism.