Research On Adaptive Low-light Image Enhancement Algorithm Based On Attention Mechanism And Feature Fusio

Low light image enhancement is an important research direction in the field of computer vision,with wide applications in real-life scenarios,especially in night-time surveillance.Images obtained in poor lighting conditions suffer from low visibility,poor contrast,color distortion,and noise.Such low light images not only severely affect human visual experience but also pose difficulties in subsequent visual task processing.Therefore,improving the quality and visual effects of low light images effectively has been an important issue of concern in computer vision research.This paper explores the illumination non-uniformity problem in low light image enhancement tasks based on deep learning techniques and improves on existing algorithms that only process channel information.The specific research content is as follows:(1)This paper proposes an adaptive low light image enhancement network algorithm based on contextual attention to study the illumination non-uniformity problem in low light images.This method first extracts initial features from the image through convolution to obtain four different contextual information features.Then,it refines the extracted features using a feature enhancement module that contains pixel and channel attention mechanisms.Finally,it inputs the learned enhancement features into an adaptive fusion adjustment module to fuse and adjust them with shallow features to restore images that conform to real lighting conditions.The experimental results indicate that compared to some classical algorithms,the proposed method is more visually appealing to the human eye in terms of subjective visual perception and has also achieved significant improvements in objective evaluation metrics such as PSNR,SSIM,and MAE.(2)Since the first work only performs feature extraction at the same scale,the brightness enhancement for objects of different sizes in the image is not significant.Based on this,this paper proposes a low-light image enhancement algorithm based on attention U-net and dualbranch residual fusion to better utilize global and local information and feature information at different scales.This method has two branches.The first branch first extracts features of different scales from the image through the U-net network,and then enhances the features using an enhancement module that contains pixel attention,and inputs the learned enhancement features into a feature fusion module to better combine local and global image information to obtain a feature map.The second branch extracts features through a residual feature extraction module,inputs the extracted features into a feature fusion module to fuse them with previous features to obtain a feature map.The two feature maps obtained by the two branches are concatenated and convolved to obtain the final enhanced image.Experimental results show that this work achieves significant improvement based on the previous work.On the real light dataset LOL,the objective evaluation metric PSNR improves by 0.56,and SSIM improves by0.072.(3)Low light image enhancement software system: Based on the proposed algorithm,this paper designs and develops a software system that implements low light image enhancement.First,the software has a concise and clear human-computer interaction interface that allows users to perform operations such as image selection,image enhancement,and image saving.Second,the software is easy to use,and users can easily learn to use it to select and enhance low light images.Finally,the enhanced images are displayed on the software interface,and users can save the enhanced results to a custom path to obtain the result image.The enhanced software in this paper satisfies the requirements of practical applications to some extent and can obtain better enhancement results.