Research On Low-Light Image Enhancement Algorithm Based On Deep Learning And Retinex Theory

As an effective information carrier,images have a wide range of applications in daily life.However,in complex lighting environments such as at night,backlighting or when the exposure time of the image capture device is insufficient,the captured images often suffer from low contrast,low brightness and lack of information.Such images not only give a poor visual experience,but also seriously affect the performance of high-level computer vision tasks,such as target detection,target recognition and object tracking.Therefore,in order to improve the visual performance and usability of low-light images,it is of great practical importance and application value to enhance them.This paper firstly introduces the basic theoretical knowledge system of low-light image enhancement and deep learning,and then introduces the traditional methods and deep learning methods in the field of low-light image enhancement,reviewing their advantages and disadvantages.Then,based on the existing methods,two low-light image enhancement algorithms are proposed,combining Retinex theory and convolutional neural networks to address the existing shortcomings.The specific research of this paper is as follows:(1)An improved low-light image enhancement algorithm based on Retinex theory is proposed to address the problems of uneven light distribution and severe noise in dark light areas of low-light images.The key to Retinex theory is to obtain accurate light information,so we decompose the low-light image into a reflection map and an illumination map using a data-driven approach.Decoupling the original space of the image into two smaller subspaces allows for better training and learning of the network on the one hand,and on the other hand allows for separate processing using the divide-and-conquer idea.The degradation is complexly distributed on the reflectance map,so we recover it under the guidance of the illumination map.The reflection network abandons the traditional denoising method and uses the Unet++ network with excellent feature extraction to process it,which not only effectively reduces the noise of the reflection map,but also preserves the texture detail information of the image more completely.For the illumination component,a full convolutional network is designed to perform adaptive brightness enhancement on the illumination map.Finally,the enhanced illumination map and the noise-reduced reflection map are combined into a clear and enhanced image.The experiments show that the method can effectively reduce the image noise,enhance the image brightness and have a good recovery of the image texture details.(2)Most existing algorithms can effectively solve the brightness enhancement problem of low-light images,but there is still much room for improvement in texture details and colour distortion.To address this condition,the paper proposes a non-uniform light image enhancement algorithm combining an attention mechanism with improved Unet based on the framework of(1).Firstly,in order to obtain a smooth illumination map and improve the accuracy of illumination information,a Unet-like network with stronger segmentation capability is designed to decompose the original image.Secondly,an improved Unet is used to process the reflectance map for solving the problem of insufficient detail enhancement and colour distortion under extreme conditions.A multi-scale residual module is introduced to replace the convolutional layers in the traditional Unet to further improve the image representation;an attention mechanism is added to the skip connection part to reduce the semantic gap between feature maps;meanwhile,a color loss function is added to address the colour distortion problem.Finally,the depth of the full convolutional network is adjusted in the illumination part and the idea of a location strategy is introduced to make it better at maintaining texture information.Experimental results show that the method in this paper has appropriate brightness enhancement,significant noise suppression,and further improvement of image details and colour information.