Research On Multi-Scene Image Enhancement Algorithn Based On GAN

Image enhancement is an important branch in the domain of digital image processing,which is widely applied in public safety,industrial production,and biomedical domains.Underwater,low illumination and foggy days are the three most commonly encountered scenes that cause image quality deterioration when people collect images.The causes of image quality deterioration in each scene are different,and the relevant enhancement technologies applied are also different.For the above three scenarios,this paper designs image enhancement algorithms based on Generative Adversarial Network(GAN)to achieve image quality improvement tasks.In underwater scenes,this paper combines attention mechanism to the GAN,constructs the UICE-GAN(Underwater Image Color Enhancement Generative Adversarial Network)based on a multi-scale channel weighted attention mechanism.On the underwater synthetic dataset,after the color cast images are enhanced by the UICEGAN,their PSNR(Peak Signal-to-Noise Ration)and SSIM(Structural Similarity)scores have doubled The proposed algorithm also obtains the best NIQE(Nature Image Quality Evaluator)score of 4.4871 on the underwater open dataset.In low-light scenes,this paper combines the Retinex theory and the GAN algorithms flexibly and proposes the LWEGAN(Low-Light Enhancement Generative Adversarial Network)lightness enhancement network.This paper also builds a new local lightness discriminator,which can judge the true or false of the average lightness of the image from a local angle,thereby enhancing the lightness of the image.On the low-light synthetic hybrid dataset and the open dataset,the LWEGAN obtains the SSIM scores of 0.93 or more,and the PSNR scores also exceed 23dB(diciBel).In dehazing scene,this paper uses the GAN to adaptively estimate the relevant parameters of the atmospheric scattering model to obtain a clear image,designs the AEG AN(Adaptive Estimation Generative Adversarial Network)dehazing network.This paper also builds a feature map fusion module,which uses a multi-level feature fusion method to improve the AEGAN performance so that the proposed AEGAN obtains the best NIQE score of 6.0144 on the real open scene dataset.