Research On Low-Light Image Enhancement Algorithms Based On Generative Adversarial Networks

Low-light image enhancement is an essential task in the computer vision field.Due to the influence of low luminance,short exposure,etc.,the captured images always suffer from a variety of degradations,including missing details,dull colors,and obvious noise.The poor imaging quality affects humans’ visual effects and makes advanced computer vision tasks difficult.Based on the generative adversarial network(GAN),this thesis proposes to improve the quality of the low-light images by increasing the brightness,recovering colors,restoring details,and reducing the noise.The main work and contributions are:(1)proposing a low-light image enhancement model(Retinex-GAN)based on the Retinex theory and GAN to address the problem of poor-quality generation by previous Retinex-based methods.Specifically,the proposed model consists of three cascaded modules:a three-channel Retinex decomposition module firstly separates the input into the luminance and reflectance components;then,a luminance adjustment module estimates the luminance adjustment parameters of the low-light image to obtain a preliminary enhancement result;finally,inspired by the adversarial learning idea,a noise refinement module further finetunes the preliminary result and generates the high-quality output.The experimental results on the LOLvlreal and LOLvl-synthetic datasets demonstrate that the proposed model achieves performance improvements compared with previous methods.Besides,an acceleration strategy is further provided to reduce computations by approximately 25%and speed the Retinex-GAN faster.(2)proposing an unsupervised low-light image enhancement model(Natural BrightenGAN,NB-GAN for short)by extracting the structural similarity and color consistency to address the insufficient robustness of supervised learning methods in practical applications.The proposed NB-GAN consists of four modules:the enhancement module,the structural similarity module,the color consistency module,and the naturalness discrimination module.Among them,the enhancement module is optimized by the other three modules.The structural similarity module establishes the structural correspondence between images of different luminance.The color consistency module supervises the color of the generated images.The naturalness discrimination module enhances the brightness of the images and ensures the realism of the results.Experimental results on five real public datasets(i.e.LIME,MEF,DICM,VV,NPE)demonstrate the superiority of the proposed model since it achieves the lowest average NIQE metric and generates the most natural results compared with previous methods.(3)proposing an exemplar-guided low-light image enhancement model(EGLLIE)to address the problems of poor generation on extremely low-light images and uncontrollable enhancement.The proposed EGLLIE employs the exemplar image as an additional input for guiding the enhancement.It consists of five sub-modules:the low-light encoder,the exemplar encoder,the decoder,the region matching module,and the attentional feature selection module.Among them,two encoders extract the features of the corresponding inputs,respectively.The region matching module learns the matching relationship between the low-light image and the exemplar image;the attentional feature selection module selects the required low-light features and exemplar features fed into the decoder module for enhancement.Besides,two exemplar datasets named pseudo-LOL and EG-office are collected to evaluate the performance of the EGLLIE model.The experimental results show that the EGLLIE improves PSNR and SSIM by about 30%and 1%on the Pseudo-LOL dataset and generates high-quality results on the extremely low-light EG-office dataset.Besides,the EGLLIE can control the degree of enhancement by adjusting the luminance of exemplars.(4)proposing a low-light image enhancement model based on the transfer connection module to address the insufficient convergence of the UNet model due to the significant difference between the encoding and decoding feature distributions.Inspired by the style transfer models,the proposed module strives to reduce their difference by transferring the style of decoding features to the encoding features.This thesis proposes to equip the UNet,Retinex-GAN,and NB-GAN models with the efficient,yet effective,transfer connection module to evaluate its performance.Experimental results on the LOL-v2 dataset show that the proposed module significantly boosts the performance with a tiny increase in parameters and computations.Especially,the PSNR and SSIM metrics of the UNet model are improved by about 22%and 5%respectively,and the LPIPS① is reduced by about 18%.