Low-light Image Enhancement Based On Multipath Residual Estimation

The computer vision system relies on images to perceive the real world,and high-quality photos bring more helpful information to the computer vision system.With the rapid development of computer vision technology,image capturing and processing technology is also constantly updated to make high-quality images easier to obtain.However,photos taken in a low-light environment are often accompanied by low brightness,low contrast,and intense noise due to insufficient ambient light,which directly affects the quality of the captured image,reducing the visual experience and affecting the implementation of subsequent high-level visual tasks.In this case,low-light image enhancement technology is promising.Low-light image enhancement is an essential branch of digital image processing technology,which aims to improve the brightness of the image taken in a low-light environment and restore the texture details and actual color of the picture.This paper discusses the existing low-light image enhancement algorithms,analyzes the problems and shortcomings of the current methods,and proposes improvement measures.The specific research methods are as follows:1.A low-light image enhancement method based on texture detail recovery is proposed.This method decomposes the low-light image enhancement into context texture and spatial detail recovery subtasks.It uses the contextual information recovery and spatial information recovery dual-branch network to focus on the image’s contextual information and spatial details through two paths while improving the brightness of the low-light image,preserving the rich semantic content and fine texture details of the original image.In addition,a new feature fusion strategy is proposed,which complements and fuses the contextual information and spatial details extracted from the dual-branch subnetwork to improve the network performance further.In the contextual information recovery sub-network,to reduce the loss of details caused by bilinear down-sampling,the complete information down-sampling method is used to supplement the details and fully retain the texture information of the image.A large number of experimental analysis shows that the method proposed in this paper can improve the brightness of low-light images,effectively reduce image noise,and produce high-quality enhanced images with vivid colors and rich textures.2.A low-light image enhancement method is proposed based on global and local illumination perception.This method uses residual learning to estimate the brightness difference between low-light and normal-light images.Based on the traditional encoding and decoding structure,this method combines the advantages of feature extraction of Transformer and CNN to model global feature dependency and local feature relationship,respectively.CNN uses a fixed-size convolution kernel for convolution,pays attention to the local feature information in the window,and improves the efficiency and generalization of image processing.Transformer has a unique self-attention mechanism,which can model the global feature information of the input sequence,making up for the limitation of the CNN receptive field.Therefore,the proposed method can balance the global and local illumination of the image and obtain high-quality images with uniform illumination distribution.Specifically,a multi-level feature extraction strategy is proposed to extract the global and local information of the input low-light image from multiple branches,which is conducive to the recovery of global illumination and local details.At the same time,a multi-level attention fusion module is designed to automatically filter the multi-level features of the previous stage output,improve the value of useful information,and compress the transmission of useless information.