Research On Low-light Image Enhancement Based On Multi-scale Retinex Network

Images captured in complex lighting conditions such as dusk,night and shadow occlusion by photography equipment usually suffer from many problems such as low brightness and color distortion.These problems make the visual effect of degraded low-light images get worse and disturb the performance of computer vision systems.Low-light image enhancement can ameliorate the quality of images effectively,provide high-quality input for advanced computer vision tasks,and increase the performance of system greatly.Therefore,the research for low-light image enhancement has great significance.This paper will focus on low-light image enhancement-related research work.Firstly,it introduces the imaging principle of the human eye,the color constancy theory,the Retinex theory and its application in image enhancement.Then,various low-light image enhancement algorithms and relevant deep learning knowledge are elaborated at length.Lastly,a novel low-light image enhancement algorithm named lightweight multi-scale Retinex network is proposed.It combines with deep learning and bases on multi-scale representation of image from multi-task and multi-scene.The network consists of several modules to fully explore and utilize the feature information.Initial illumination is estimated by extraction module from different scales independently,then it is fused across scales guided by progressive fusion strategy in fusion module.Finally,the fusion illumination is further integrated and smoothed by restoration module and get refined illumination,the effect of external illumination is eliminated on the basis of Retinex theory,and obtain the enhancement image.The network is trained using various low-light datasets to make it robust to low-light images from multi-scenes.Then the well-trained network is tested on artificial and natural low-light datasets respectively.Extensive experiments with other advanced enhancement algorithms demonstrate our proposed algorithm can enhance low-light images from multiple scenes and improve the quality of images effectively.Moreover,the complexity of our algorithm is low and has practical value for parallel computing of the network.