Research On Low-light Image Enhancement Based On Densely Convolutional Networks

In social life,computer vision technology is used in many fields,and it has been widely used in many fields such as video surveillance,robot vision,and drone reconnaissance.At present,many scholars have carried out a lot of research on normal-light images,but there are few researches on low-light images.Images captured in scenes with insufficient overall illumination or uneven local illumination suffer from low brightness,insufficient contrast,and severe loss of detail information,which make such images unable to meet the expected requirements and seriously affect the subsequent applications of such images.In this paper,we focus on low-light image enhancement to improve the quality of low-light images with richer detail information by studying image enhancement techniques.Low-light scenes can be divided into global insufficient illumination and local uneven illumination.In this paper,using dense connection network as the key technology,the following two image enhancement methods are proposed respectively for the above different scenes.1)In order to improve the quality of images acquired under global illumination deficit conditions,it proposes a low-light image enhancement method based on Attention Residual Dense-Generative Adversarial Networks(ARD-GAN)is proposed by combining generative adversarial networks with residual dense structures and introducing attention mechanism.First,the method generates a global exposure attention map in the Global Illumination Estimation Module(GIEM)to guide the subsequent modules for better illumination enhancement;second,it uses the Convolution and Residual Module(CRM)and Channel attention Residual Dense Module(CARDM)to extract shallow and deep features respectively,and fuse the extracted features at different levels to obtain better detail information;then,combining dense connectivity with batch normalization on the basis of CARDM to suppress noise,Dn CNN supervised generation network is used to achieve better noise reduction;finally,the loss function is improved and combined with multiple loss functions as a joint loss function to make better color reproduction of the enhanced image.The experimental results show that ARD-GAN effectively enhances low-light images and has better performance of enhanced image details,noise reduction and color reproduction.2)For low-illumination scenes with uneven local illumination,this chapter proposes an a low-light image enhancement network for solving the problem of uneven local illumination in low-light images is proposed by combining a densely connected convolutional network,an error feedback mechanism,an attention mechanism and a Transformer,called a multiscale error feedback network.First,the low-light image is fed into the error feedback coding module to compress and encode the image information,where the error feedback mechanism can play the role of complementing the high-frequency information;secondly,the encoded information is fed into the feature integration module to model the global context of the encoded information by using the features of Transformer;finally,the modeled information is fed into the error feedback decoding module for feature decoding to get the enhanced image.Experiments show that MSEFN with the Local uneven illumination(LUI)dataset constructed in this paper for training can give different degrees of illumination enhancement according to different locations of low-light images,and it has certain advantages in visual effects and evaluation indexes.