Low-light Image Enhancement And Its Structural Similarity Enhancement Based On Deep Learning

Images captured under low-light conditions cause people’s sensory vision into a blind area,and the essential attributes of objects are hidden in the dark.These defects are hidden in the dark due to camera defects and quality that cause noise and color distortion.Firstly,low-light images seriously affect human visual perception;secondly,due to the development of computer vision,there are increasingly high requirements for sample pretreatment,and low-quality images will affect its performance.Thus,research on low-light image enhancement is particularly important,the image denoising,compensate for missing information,distortion,structure similarity,the enhancement and improvement on these problems will provides high quality data samples to target recognition and tracking,video enhancement,urban city security and intelligence development work,the study has important research significance and practical value.There are many excellent algorithms that can adjust the brightness of low-light region,but they are not enough to realize the function of denoising and similarity restoration of inner attributes of objects.This paper first introduces the theoretical support of low-light image enhancement and neural network,then introduces the traditional methods of low-light image and deep learning methods,and analyzes their advantages and disadvantages.On this basis,two methods for low-light image enhancement are proposed.The specific research contents of this paper are as follows:(1)To solve the problems of excessive enhancement and inconspicuous illumination adjustment of low-light images,we propose the Attention Mask-guided low-light image enhancement algorithm-LLENET,inspired by the Retinex model.This algorithm is designed with two main parts: decomposition network and enhancement network.By decomposing the image into a light map carrying the illumination information and a reflection map carrying the inner information of the object,the image is reconstructed after enhancing them separately,so that the enhanced image is more realistic and avoids the lack of information on the inner properties of the image due to some existing methods enhance the image only on the light map.In the enhancement part,we divide it into image restoration and adjustment.In the restoration part,we need a model with more perfect image information retention and denoising ability,so we give up the traditional interpolation method for restoration and use U-NET instead,and add Attention Mask as guidance before U-NET enhancement channel.The nature of the network is still end-to-end trainable,and the low-light dataset LOL was used in the experiment.Through experiments and comparative analysis,this method has better ability of denoising and improving image quality,and the overall image brightness adjustment has more advantages than the comparison methods.(2)Some of the existing algorithms can adjust the brightness of low light regions,but they are slightly inadequate in achieving the functions of denoising,similarity restoration and image information retention of the intrinsic properties of objects.Inspired by the working principle of human retina,the image is decomposed into two modules by the designed decomposition network,named the illumination component(reflecting the intensity of illumination)and the reflection component(reflecting the intrinsic properties of objects),and then transports both to the enhancement network designed in this paper for enhancement,which is guided by the modified Attention Mask module for component guidance,and the enhancement is completed by the U-NET network structure.In the U-NET structure network,a RISS module is designed in this paper,which is used to effectively reduce the loss of image feature information and better retain the authenticity of the image,so as to achieve the structural similarity of reconstructed images.For illumination adjustment,an improved convolutional network is used to ensure the position relationship between illumination feature information and the original image.In this paper,the network is trained on image dataset(LOL)with different illumination levels,and it is proved through experiments that the network can adapt to different intensity of illumination images to restore them.Later,by comparing with other methods(traditional methods and deep learning methods)on different datasets subjectively and objectively,the method can not only realize the functions of low-light image enhancement,but also the functions of color information restoration,image detail restoration,image structure similarity enhancement and image denoising are better improved.