Attention-Driven Photo Improvement Architecture for Low-Light Spaceborne Image Assessment

Abstract

Night-time satellite imagery is very vital in areas such as environmental monitoring, urban planning and disaster management. However, there is a problem in performing remote sensing mainly due to issues of change in illuminance level where structures in images may become blurred and noisy. Most traditional techniques of image enhancement were unable to produce good results uniformly and they also failed to balance the important features of the image under different illuminations. In order to overcome these limitations, the Attention-Driven Photo Improvement Architecture for Low Light Spaceborne Image Assessment is designed. The underlying concept of this framework is based on self-attention mechanisms and multi-scale feature extraction to increase image contrast and visibility. It also applies architectural features such as self-supervised learning, contrastive learning and changes proposed in Retinex-recurrent to enhance the image texture but not the structure. The comprehensive experimental results on several test databases also confirm that the current approach achieve higher PSNR, SSIM, and FSIM values than the deep learning methods. Moreover, through bringing down the computational cost to 30%; it draws path for enhanced, real time analysis of satellite images to open up new frontiers of feasible research in the field of remote sensing.