BCNN Design For Visual Prosthetic Image Classification

A visual prosthesis is an embedded system that constructs an image by controlling electrodes.It constructs a lattice by implanting electrodes in the human body to form an image,and then sends the constructed image to the blind visual system to restore part of its vision.Since the number of electrodes implanted in the human body is limited,the image recognition of the electrode composition is generally low,and the use effect of the visual prosthesis is not satisfactory,and there are still many difficulties.Therefore,this paper proposes to use the classification neural network to assist the visual prosthesis work,thereby reducing the difficulty of using the visual prosthesis.In order to reduce the difficulty of classification neural network in visual prosthesis embedded system,the main work of this paper is to improve the binary neural network algorithm and parameter compression.In this paper,three algorithms of AlexNet-Binarized,InceptionNet-Binarized and IBNet are proposed for the shortcomings of binarization algorithm BinaryNet.IBNet uses double identity mapping method to improve the accuracy of the network and uses two hyperparameters to control the identity mapping.The degree of performance of the three algorithms on the benchmark test set CIFAR-10 was tested.The test results show that IBNet has the best overall performance.Finally,this paper compares the performance of IBNet and BinaryNet algorithms on the self-made indoor environment dataset,and compresses the weight parameter,and tests the speed of BinaryNet and IBNet processing 32x32 images.The accuracy of AlexNet-Binarized in CIFAR-10 reached 75%,the accuracy of InceptionNet-Binarized in CIFAR-10 reached 83%,and the accuracy of IBNet in CIFAR-10 reached 91%,while BinaryNet was in CIFAR.The accuracy rate on the-10 is only 88.9%.IBNet's accuracy rate in the indoor environment classification data set reached 98%,BinaryNet's accuracy in the indoor environment data set was only 88%,and the model's compressed IBNet parameter quantity was only about 800,000,and the BinaryNet algorithm's parameter quantity exceeded 10 million.Finally,the FPS of the IBNet processed image reached 45,and the FPS of BinaryNet processing the same size image was only 18,which showed that IBNet has better performance than BinaryNet.The main purpose of this paper is to provide an algorithmic reference for visual prosthetic embedded system implementation.