Research On Denoising And Enhancement Of Low Light Image Based On Convolutional Neural Network

In recent years,image processing technology based on deep learning has been continuously developed along with continuous research on deep learning.Based on the research of low illumination image denoising and enhancement,this paper focuses on the deep learning technology based on convolution neural network.Analyzed the traditional image denoising and enhancement algorithms.the popular classical neural network models in recent years are studied and compared.In this paper,the convolutional neural network model U-Net and CAN are finally selected for low illumination image processing.The emphasis includes pre-processing of low illumination image,semantic segmentation module of convolutional neural network and super-resolution reconstruction module in later stage.The main work of this paper is as follows:First of all,the pre-processing of low illumination image:In this paper,we selected the open source low illumination image dataset SID(See-in-the-Dark),which contains low illumination short exposure photos and corresponding real value long time exposure photos,and the data sets include low illumination short exposure photos and corresponding real value long time exposure photos Because the low illumination short exposure photo is almost black,the black horizontal extraction and magnification are carried out in the early stage to improve the brightness.Secondly,the pixel semantic segmentation of the pre-processed images is carried out,and the convolution network U-Net and CAN are used to do the research.According to the different characteristics of the network structure,in the aspect of network optimization,U-Net adopts the adaptive learning rate algorithm Adam to update the parameters,and the CAN network adopts the batch standardization optimization strategy.Finally,in order to improve the resolution of the image,ESPCN super-division reconstruction is used to improve the resolution of the image.At the same time,the peak signal-to-noise ratio(PSNR,Peak Signal to Noise Ratio)and image self-similarity(SSIM,Structural Similarity Index)are used to evaluate the image quality.This article carries on the experiment based on the Ubuntu system Tensorflow environment,carries on the analysis and appraisal to the experiment result.The experimental results show that compared with the traditional image processing thread and denoisingalgorithm,the end-to-end image processing thread based on convolution neural network has the advantages of simple implementation and higher image quality.Image processing based on U-Net network model has higher peak signal-to-noise ratio(PSNR)and has obvious advantages for very low illumination image processing.