Low-light Image Quality Enhancement And Application Based On Deep Learning

Images are the basis of human vision and an indispensable part of the development of human society.Due to the inevitability of low-light environments,acquired images often have problems such as low brightness,high noise and lack of details,which will bring many difficulties to subsequent image processing,such as target detection,industrial production,remote sensing and monitoring.Low-light image enhancement has become an urgent problem to be solved.This paper proposes two effective methods to solve this problem,as follows:(1)To solve the phenomenon of uneven brightness and color cast that are easy to appear in the process of low-light image quality enhancement,this paper designs a brightness enhancement network model.To better preserve the hue and saturation information of the image,this paper converts the image from RGB to HSV color space,and takes out the luminance channel separately for processing.Based on the residual network and the principle of iterative enhancement,this paper fuses feature maps of different brightness to obtain a brightened image.The model is applied to daily life scenes,and images with better visual effects are obtained.(2)To solve the phenomenon of missing details and blurred edges that are easy to occur in the process of low-light image quality enhancement,this paper designs a continuously updated connection network model.Based on the U-Net framework and the idea of iterative enhancement,this paper fuses features of different scales to obtain more global and local information.Based on the channel attention idea,this paper assigns appropriate weights to each channel to obtain more details.This paper also performs residual weighting on the obtained images of different scales to obtain more accurate color information,so as to obtain clear and beautiful brightened images.The model is applied to the visual synchronous localization.After the brightened image is subjected to the synchronous localization experiment,the results show that the AUC(Area Under Curve)is improved by 0.04,which effectively increases the accuracy of the positioning system.This paper evaluates two proposed methods on a common dataset and compares them with the optimal values of various existing methods in three metrics: SSIM,PSNR,and NIQE.The results show that both methods proposed in this paper achieve the best results.