Skynet-image Enhancement Application Research Based On Deep Learning

With the increasing popularity of digital informatization in today's society,people have to extract the interesting information from the massive information of the outside world every day.Images are one of the main ways that people transmit,obtain and process information.The imaging quality of an image is often affected by the external conditions.For example,it can only illuminate a local area on the surface of the object when the light becomes weak at night or bad weather.The local area information is obscured by the darkness,so the images collected by the "Skynet" system are not uniformly exposed which will seriously affect the algorithm's processing of the images collected by Skynet.Low-light image enhancement based on deep learning is a transformation of an image processing system that can restore low-light images to normal-illuminated images through learning.It is an application of artificial intelligence technology.Under the application of this technology,it not only greatly reduces the cost compared with higher hardware equipment performance to improve image quality,but also can intelligently process large-scale low-quality images.This technology is very important for advanced image processing tasks such as target detection and image recognition.Aiming at the problem of strong noise interference in low-light images,this paper proposes a deep learning method for low-light image enhancement with joint decomposition and denoising.First,based on the structure of the U-net model,a joint decomposition and denoising U-net network(JDEU)which decomposes the low-light image into reflectance and illumination containing noise is constructed by adjusting the operation of the sampling layer.After approximating the noise-free reference image to the reference reflectance,the reflectance consistency restores the color and overall structure of the desired reflectance While the structural similarity can speed up the suppression of the noise from the reflectance.Then,the multi-scale brightness enhancement network increases the light intensity of the initial illumination.Finally,the denoised reflectance and the enhanced illumination are merged into an enhanced image.A large number of experiments show that the proposed method not only has a significant denoising effect,but also retains structural detail information well.It is difficult to obtain the paired reference images in practical applications,which seriously affects the enhancement ability of the low-light image enhancement technology in different scenes.This paper proposes a color variation minimization Retinex decomposition and enhancement based on a multi-branch decomposition network(Cvm D-net)to enhance low-light images without reference.Firstly,an input constant feature prior mechanism(ICFP)is proposed according to the inherent properties of the image,which uses the average intensity feature of the low-light image to approximate the reference image and guides the decomposition and enhancement process.Secondly,a multi-branch decomposition network learns the different dimensional characteristics of reflectance and illumination.The network simultaneously predicts the reflectance and illumination of the low-light images through the different sub-networks,and further constrain the color and noise in the reflectance based on the minimization of the color variation between the input image and the reflectance.Finally,the predicted initial illumination is enhanced by the average intensity of the input image.A large number of experiments show that the proposed method can improve the quality of the original image well without reference image,and shows good results in many real scenes.The above methods are applied to the Skynet video surveillance image processing platform,and the night surveillance images are captured to test the image enhancement system in the platform.Experiments show that the methods of this paper are very effective in improving the quality of night images collected by the "Skynet" system,and provides convenience for public security night investigations.