| Along with the popularity of photographic devices,images are relevant to almost all human activities.Under low-light conditions,images captured by photographic devices suffer from low contrast,poor visibility and high noise.It not only affects the human visual perception ability,but also poses a great challenge to computer vision systems.How to improve the visibility of low-light images with better color,richer detail features and lower noise is a classical problem in the current image processing field.Although a series of low-light image enhancement algorithms have emerged in recent years,the enhanced images of these algorithms still suffer from chromatic aberrations,artifacts and loss of details.This paper presents a in-depth study of this task of low-light image enhancement.First,this paper focuses on focusing on dark areas to enhance the overall brightness of the image and increase the visibility of the enhanced image;second,to prevent the problem of low saturation,high noise and color bias in the enhanced image,two efficient algorithms are designed in this paper.To solve this series of problems,this paper proposes corresponding solutions based on the end-to-end model and considering the advantages and disadvantages of existing methods.The work in this paper is summarized in two main aspects:(1)A Low-light image enhancement via multi-stage feature fusion network is proposed.The process is mainly divided into two aspects,multistage feature extraction and multi-stage feature fusion.In the multi-stage feature extraction,the low-light images and the illumination images are firstly pre-processed in order for the model to focus more on the dark light region of the image and reduce the burden of the model operation,and then the feature extraction process is divided into three stages,and the advantages of the spatial weighted residual channel attention block in UNet module and the non-local sparse block in single scale module are combined to extract the accurate feature information.In the multi-stage feature fusion,firstly,the key feature information of the previous stage is passed to the next stage by the spatial weighted residual channel attention block and fused with the features of that stage,which effectively solves the problem of repeatedly passing a large amount of non-key information in different stages;secondly,a fusion attention mechanism is designed,which can provide effective supervision for each stage to control the passing of a small amount of key feature information.Finally,the proposed method is proved to be more competitive through extensive experiments.(2)A low-light image enhancement via illumination prior-aided network is proposed.The proposed network is designed based on Retinex theory.First,an illumination prior guidance strategy is designed to precisely control the exposure of the image to avoid the problem of underexposure of the image.Secondly,a light information extraction module combined with the efficient channel attention in residual block is designed in order to avoid color bias and low saturation problems.In addition,an image enhancement module based on wavelet transform is designed to extract multi-scale feature information accurately.Finally,an interactive feature fusion module is designed to better apply the model to complex scenes with light.The superiority of the proposed algorithm is demonstrated by extensive experiments. |
