Low-light Image Enhancement

With the increasing popularity of shooting equipment,images play an irreplaceable role in the development of human society.Images can not only record daily life,but also have a wide range of applications in industry,so it is essential to obtain images with good visual effects.In the actual shooting process,there are many external or internal factors that will affect the quality of image shooting,and the lighting environment is an important factor.Images acquired under low light conditions have the disadvantages of low brightness,high noise and missing of detailed texture,which can not meet the requirements of user.Low-light environments are inevitable,and while you can improve the quality of your images with hardware,the cost is prohibitively high.At present,enhancing the low-light images obtained is a popular research direction.This article mainly uses deep learning to enhance low-light images,Specific work includes:The first chapter is mainly about the background of low-light image enhancement,and the second chapter gives a more detailed description of image enhancement and deep learning.In the third chapter,combined U-net and residual network,a front-end dense neural network is proposed to enhance low-light images.The network can effectively use the original information of the input image.The neural network adds preprocessing before enhancing the low-light image.By adding a fixed value to the pixel value of the low-light image,the brightness of the low-light image is increased to the brightness of the image taken under normal light.Preprocessing can effectively improve the quality of the input image and improve the training results of the neural network.In the fourth chapter,the Retinex theory is used to construct an enhancement network.The network is divided into two parts.The first sub-network adopts the method proposed in Chapter 3 to obtain a saturated image,but the texture edge is still lacking.The second part is the decomposition network which can obtain the reflection of the image,and the reflection map has useful edge texture information.The two networks use the same design concept,try to use the original image information to obtain more realistic results.The output images of these two networks are input into the guided filter together to obtain a further enhanced output image.After experimental comparison and verification,the method proposed in this paper can effectively enhance low-light images to obtain satisfactory results.