Research On Neural Network Of Non-linear Image Distortion Correction And Identification Technology

This paper analysis the basic principle of artificial neural network, focuson the neural network’s application on two aspects:nonlinear image distortioncorrection and image identification. In the nonlinear distortion correction,we study aBP neural network model based on the Levenberg-Marquardt algorithm to correctnonlinear image distortion;In the section of image recognition,we establish a netwokwhich is added momentum algorithm to get Arabia digital image identification.Forthe larger image distortion,which is difficult to identify,a cascaded neural networktechnology is proposed.The first neural network is used to get image distortioncorrection,Then with the help of second neural network we can get imagerecognition.The main research contents are described as follows:One BP neural network model based on the Levenberg-Marquardt algorithm isresearched.For the nonlinear image distortion,we select the characteristic pixelsproperly to train neural network model.After trained the neural network can achievethe image distortion correction.The experimental results show that, the networkmodel’s effect on image distortion correction is ideal.And the corrected image erroris less than0.8pixel coordinate value.Draw a VC development environment standard images and noise images basedon VC development environment,then we finish the digital character recognitionusing neural netwok.For the seriously distortioned image which is difficult toidentify,we present a cascade neural network model.The first is BP network,which isbased on Levenberg-Marquardt algorithm using MATLAB neural netwok toolboxand the second is a netwok which is added momentum algorithm based on the VCdevelopment environment.We can use this cascade neural network to finish imagecorrection firstly,then finish image identification using the second neuralnetwork.This can improve the accuracy of image recognition greatly.Theexperimental results show that,the rate of digital image’s recognition whose distortion rate is under twenty percent is up to one hundred percent for trainedcascade neural network.A method by optimizing the alignment mark pattern with a proper neuralnetwork to improve the lithography alignment accuracy is proposed.And we use anexperiment of using neural network to realize the correction of self-designed " cross" alignment mark to verify the theoretical feasibility.