Research On Low-Light Image Enhancement Network Based On Multiscale Feature Enhancement

Photographs taken under underexposed conditions usually suffer from certain degradations that are difficult to model and predict,and these degradations make the images suffer from insufficient brightness,low contrast,noise and color,and such low-light images are not conducive to human observation and subsequent applications in advanced vision tasks.To address these problems,low-light image enhancement techniques have emerged as the underlying task in the field of computer vision to obtain brighter and clearer images through the use of image enhancement techniques.This task is of great practical importance in vision fields such as night surveillance,driverless cars and object detection.In the past decades,many researchers have worked on the development of low-light image enhancement techniques.Among them,low-light image enhancement methods using deep learning as a framework stand out among many types of methods,which have strong generalization and better performance,but there are still some problems that have not been solved,such as strong noise,low contrast and color bias.To compensate for these deficiencies,we propose two new low-light image enhancement schemes,which are as follows(1)To address the shortcomings such as insufficient enhancement of dark areas,strong noise,and blurring that still exist in enhanced images.In this paper,we propose a Progressive Enhancement Network Based on Detail Supplementation for Low-Light Image(PEN-DS),which is implemented by building two modules: Image Preprocessing Module(IPM)and Progressive Image Enhancement Module(PIEM).The IPM constructs an image pyramid structure to obtain low-light images and detailed images at different scales.images at different scales.In addition,the method employs a multi-supervised approach for enhanced images at different scales in order to better train the network.Extensive experimental results show that the method outperforms current research methods in terms of visual perception and objective evaluation.(2)To address the problems of overexposure and color distortion of enhanced images,this paper proposes a Low-Light Image Enhancement Network Based on Multi-Scale Feature Complementation(LIEN-MFC),which is a multiscale supervised U-shaped codec image network.In the encoder,four feature extraction branches are constructed to extract the features of low-light images at different scales.In the decoder,a Feature Supplementary Fusion Module(FSFM)is proposed to complement and integrate the features from different branches of the encoder and decoder in order to ensure the integrity of the learned features at each scale.In addition,a Feature Restoration Module(FRM)and an Image Reconstruction Module(IRM)are constructed in each branch to reconstruct the recovered features and output the enhanced images.In order to better train the network,a joint loss function is defined,in which a saturation loss term and a confrontation loss term are designed to ensure that the enhanced results better meet the visual properties of the human eye.Extensive experiments show that the method outperforms the current state-of-the-art methods both subjectively and objectively.