Research On Low-light Image Enhancement Based On Generative Adversarial Networks

Image data is widely used in various fields such as computer vision,security monitoring,product inspection,human-computer interaction,and medical applications.However,in real life,there are many scenes with low light intensity,dim background light,and backlighting.At this time,the acquired image has problems such as insufficient brightness,low contrast,and loss of a lot of detailed information.The image quality of such low-light images is poor and cannot meet people's expectations.At the same time,the application value of the images is greatly reduced.This subject studies the algorithms about the low-light image enhancement.It aims to improve the quality of low-light images by researching some technical methods and methods to restore the scene information of the low-quality images,so as to obtain natural and clear images with complete details,structural information.In this paper,the following three methods are proposed for low-light image enhancement using the adversarial network as the key technology:In this paper,the problem of low-light image enhancement is transformed into the problem of transforming from low-light images to normal-light images.We used the condition-generating adversarial network structure to propose a low-light image enhancement method based on U-Net generation confrontation network,which is called LLEGAN.This method uses the U-Net skep structure to retain the image informations to improve the generation network,which improves the performance of the illumination enhancement model and makes it better in the repair of color and details.It uses the characteristics of different loss functions to improve the total loss function of the network,so that the network can obtain an the enhanced-light image with better visual and objective evaluation.Experiments verify that the LLEGAN method can effectively achieve low-illumination image enhancement,and obtain better illumination enhancement results than some classic low-light enhancement algorithms.Aiming at the problems of unstable training of generative adversarial network models and the collapse of existing models,this paper uses the conditional variational auto-encoding-generative adversarial network(CVAE-GAN)structure to construct models.By combining the advantages of c VAE and c GAN,this paper proposes a low-light image enhancement method based on CVAE-GAN,which is called c VAE-GAN-LLE.First,c VAE-GAN-LLE learns the latent state representation of low-light images through c VAE,then learns the mapping from low-light image distribution to normal-light image distribution through c GAN,and finally enhances the quality of low-light enhanced images through c GAN's counter-training.The experimental results show that the addition of the VAE makes the model more stable,and at the same time the resulting image quality is better.This paper combines the generation adversarial network with the brightness and color perception model in Retinex theory,and introduces a dense convolution block and attention mechanism module.Finally we proposes a low-light image enhancement network based on Retinex and Attention,called MDAR-GAN.First,MDAR-GAN uses multi-scale convolution to perform preliminary feature extraction on the low-light images.Secondly we uses a dense convolution blocks to perform further feature extraction on preliminary features.And then the attention mechanism is used to improve the features to estimate the ambient light and noise attention map components.Finally,the reflection component in the low-light image was obtained according to Retinex theory,that is the enhanced image.Experiments show that MDAR-GAN can effectively improve the brightness,contrast,and color of low-illumination images,and has certain advantages in evaluation indicators compared to other algorithms.