Research On Lightweight Retinal Coding Models

With the increasing demand for computer vision in mobile terminals such as mobile phones,how to efficiently encode visual signals has attracted growing attention.As a kind of model for coding visual signals,retinal coding models encode visual stimuli into a series of spike sequences by non-linear computation to simulate the characteristics of retinal cells.There are a bunch of retinal coding models,but the existing models face some problems.For example,simple models show poor fitting effects,while complex models have massive computation.Therefore,to target the above problems,this thesis focuses on studying the lightweight of retinal coding models,and the main content is as follows:(1)Firstly,since the spike-triggered average method needs a large number of samples to obtain cell receptive fields,this thesis first proposes a spatiotemporal filter based method to reduce the required samples.Secondly,aiming at the problems of large parameters and high computational complexity of existing retinal coding models,a lightweight retinal coding model based on depthwise separable convolution is proposed.The model consists of two depthwise separable convolution modules and one dense connection layer,each depthwise separable convolution module containing two depthwise separable convolution layers.Compared with the retinal coding model based on linear nonlinearity,the retinal coding model based on generalized linear and the retinal coding model based on convolutional neural network,the proposed lightweight retinal coding model based on depthwise separable convolution has fewer model parameters,stable training process,and high prediction accuracy.(2)Because of the weak robustness of single retinal coding models,two retinal coding integration models are proposed in this thesis.Firstly,the heterogeneous integration model is built by Bagging algorithm,and the experimental results show that the Bagging integration model has good robustness.Secondly,this thesis improves the strategy of updating sample weights in Adaboost R2 algorithm and proposes an Ensemble algorithm Adaboost Poisson suitable for retina coding model,which is applied to construct a homogeneous model.The experimental results show that the Adaboost integration model can enhance accuracy and robustness.In summary,this thesis focuses on the problem of large samples facing traditional methods and proposes a method for extracting cell receptive field.The experimental results show that the proposed method can obtain cell receptive fields with a small number of samples.Because typical retinal coding models suffer from many parameters and high computational complexity,a lightweight retinal coding model based on depthwise separable convolution is then developed.The experimental results show that the lightweight retinal coding model based on depthwise separable convolution has the advantages of high accuracy and stable training.Finally,two retinal coding integration models are designed to improve weak robustness of single retinal coding models,and experimental results show that the Ensemble models have strong robustness.