Low-light Image Enhancement Based On Generative Adversarial Network

Images are an important way to obtain information,and images are often captured with low brightness,blur,artifacts and different types of noise due to conditions such as low-light environment,professional level of the shooter and hardware equipment.In order to improve the image quality and content recognition,low-light image enhancement algorithm has always been a hot topic in the field of computer vision frontier research.When the low-light image is enhanced,many factors will interfere with the results,which leads to the image enhancement effect is not ideal,which seriously affects the image perception.This paper first describes the history,background and significance of low-light image enhancement in detail,introduces the principle of the classical low-light image enhancement algorithm,analyzes and understands all kinds of low-light image enhancement technologies comprehensively,draws lessons from the frontier achievements of the relevant fields such as super-resolution and fuzzy image enhancement,and puts fo rward a new idea for low-light image enhancement.The main research contents of this paper are as follows.Traditional low-light image enhancement algorithms generally only consider enhancing image contrast,and do not consider enough for noise suppression,image details,etc.Some low-light enhancement networks based on deep learning do not restore enough details of low-light images,and image blurring,color deviation and artifacts occur for low-light image enhancement with low luminance.In this paper,considering the characteristics of low-light images,we propose an enhancement model of low-light images based on generative adversarial networks,using U-Net as the generator,and Patch GAN to discriminate the truth of enhanced images and normal light images,with a loss function consisting of multiple losses,including adversarial loss,mean square error loss,VGG loss and color loss,VGG loss will make the generated images rich in high-frequency detail information,and the color loss responds to the problem of color distortion of the enhanced image.The model proposed in this paper is compared with some classical low-light image algorithms such as BIMEF,LIME,Retinex-Net and Kin D for experiments,and using the image quality evaluation standards NIQE,SSIM and PSNR to analyze the experimental results,the experimental results show that the method in this paper has a significant effect on the enhancement of low-light image quality,restores the clear details of the low-light image,and makes the image color n atural and texture obvious.For the U-Net generator network with too deep layers and too large model,which increases the difficulty of model training,this paper also proposes to use U-Net++,U-Net 3+ and Attention U-Net,to perform network optimization o f the generator,and these networks can fully extract the texture,color and other features of low-light images on the basis of reducing the sampling layers.The experimental effects of various algorithms on low-light images with different luminance are an alyzed,and by comparing the enhanced images and various image quality evaluation indexes,it is found that the model designed in this paper is more effective in improving the quality of low-light images with lower luminance.Through extensive experiments,it is shown that the improved generator network not only optimizes the model parameters,simplifies the network training difficulty,enhances the operation efficiency,improves the viewing quality of low-light images,and outperforms other low-light image enhancement algorithms in the comparison experiments.