The Research On Low-light Image Enhancement Method Based On Deep Convolutional Neural Network

Low-light image enhancement has an important role in social reality scenarios such as real-time surveillance,autonomous driving,and computational photography,and a rational analysis and application of low-light images is of great research value.Low luminance contrast,loss of details,severe noise and color distortion,and uneven illumination are problems in low-light images.Current excellent low-light image enhancement algorithms can effectively improve the brightness and contrast,but the enhanced images suffer from problems such as the need to balance noise reduction and detail recovery,color distortion,uneven illumination among the localities,and low perceptual quality.In addition,there is still a lack of real paired datasets,leading to the problem that current models trained based on paired datasets have weak generalization ability.Current models based on semi-supervised learning are trained using paired and unpaired datasets,but suffer from weak enhancement performance.The above problems lead to increased difficulty in the research of low-light image enhancement algorithms.Therefore,this paper addresses the main difficulties in low-light image enhancement tasks and combines the characteristics of low-light images to study low-light image enhancement methods based on deep convolutional neural networks,aiming to deeply explore the potential value of low-light images and provide technical support for deep analysis of low-light images.The main work of this paper is as follows:(1)For the problems that need to balance noise reduction and detail recovery,color distortion,uneven illumination between each localization and low perceptual quality in low-light image enhancement tasks,a low-light image enhancement model based on improved Kin D-LL-Kin D is proposed.first,a reflectance recovery network based on MF-Res UNet is proposed.on the one hand,a VGG network is used to extracts multi-scale features to guide the noise reduction process in order to solve the problem of balanced noise reduction and detail recovery;on the other hand,the color-related features are enhanced by the residual module and color loss to solve the color distortion problem.Secondly,the illumination adjustment network based on improved Dense Net is proposed to solve the problem of uneven illumination among localities by deepening the network and increasing the proportion of shallow features.Finally,the perceptual strengthening network is constructed,and the global features of the image are enhanced using null convolution and feature loss to solve the problem of low perceptual quality.Experimental results show that LL-Kin D performs well in low-light image enhancement tasks.(2)A semi-supervised low-light image enhancement algorithm-SLIER based on Retinex theory is proposed to address the current problems of lack of real paired datasets and weak performance of model enhancement based on semi-supervised learning.on the one hand,semi-supervised learning is combined with the low-light image enhancement task to solve the problem of lack of real paired datasets.On the other hand,the problem of weak enhancement performance of the semi-supervised model is addressed from the following three perspectives.First,a three-branch decomposition network is proposed with an improved U-Net network structure to reduce feature loss in dark areas by self-calibrating the convolutional residuals,and then adding attention gates to filter redundant noise;second,a degradation distribution loss is proposed based on the law of low-light image degradation distribution to improve the network’s ability to eliminate various degradations,and then a reflectance detail loss is proposed in combination with a fractional-order differential mask to recover reflectance texture details of the image.Finally,to solve the problem of low perceptual quality of reflectance images,an unsupervised learning-based perceptual reinforcement network is constructed and trained with a large-scale unpaired data set to finally generate high-quality images with natural illumination and texture.The experimental results fully demonstrate the excellent performance of SLIER in low-light image enhancement tasks.