Research Of Low-light Image Enhancement Based On Enlightenment Generative Adversarial Network

With the rapid development of science and technology,image data is increasing day by day in real life,but the real environment is complicated,and the lighting conditions are not good in some special situations,resulting in unsatisfactory brightness of the image.Therefore,it is necessary to use low-light image enhancement technology to further process the image to obtain better features and visual effects.Existing low-light image enhancement technologies generally use traditional image processing meth ods.Many of these traditional methods are enhanced for specific aspects,which are prone to excessive enhancement,resulting in serious color cast and loss of details.In recent years,methods based on deep learning can learn many details of image process ing,thus solving the shortcomings of traditional methods.However,the existing deep learning models use data sets in specific scenarios,and the generalization ability of the model is not strong.Therefore,this article takes the improvement of the gener alization ability of the deep learning model as the starting point,combined with the generation of confrontation networks to study the low-light image enhancement technology.This article mainly optimizes and improves the model based on EnlightenGAN to im prove the generalization ability and enhancement effect of the model.The main improvements in this article are as follows:(1)When generating images for the EnlightenGAN model,only the global attention is considered in the brightness part,and the local attention is not considered.Therefore,this paper proposes a global and local attention model EnlightenGAN v1 based on EnlightenGAN.Through the overall brightness and local brightness of the image The value of is weighted average,the average brightness attention is used to replace the previous overall brightness attention,and the attention function of EnlightenGAN is improved,instead of using simple 1-I,use a more appropriate sigmoid(1-I)In this way,the attention weight of the image brightness can be reflected more accurately,and the model effect can be better improved.By designing different comparative experiments,better results have been achieved in the Natural Image Evaluator(NIQE),and the best evaluation has also been obtained in user research experiments.Compared with CycleGAN and EnlightenGAN,the effect is significantly improved.(2)The generation network U-net for the EnlightenGAN v1 model does not consider the important issues of both shallow and deep features in the feature extract ion stage.The generation network uses U-net++ to replace the original U-net,because U-net++ can effectively extract Shallow features allow both shallow features and deep features to play their advantages together.In addition,this paper combines the los s functions and combines the advantages of different loss functions to allow the generation of confrontation networks to capture more texture features of low-light images,learn image textures,and better achieve low-light image enhancement.In this paper,we conducted experiments on screening suitable data from data sets such as lol.From the perspective of experimental intuitive effects and natural image evaluators,the effects of EnlightenGAN and EnlightenGAN v1 are significantly improved.