A Refined Zero DCE Framework For Enhancing and Interpreting Lunar Permanently Shadowed Images

Abstract

This study outlines a system driven by deep learning technologies to improve lunar imaging at low light levels to make them more useful for scientific analyses. In particular, lunar images obtained from Permanently Shadowed Regions (PSRs) exhibit very low light levels, excessive noise, and a loss of detailed features, making them extremely difficult to analyze. The proposed solution incorporates the Zero-DCE model, which establishes brightness and contrast adjustments without the need for paired training data. It will also employ Tri-Curve enhancements along with a convolutional neural network (CNN)-based refinement module to achieve improved feature preservation and reduced noise in lunar imagery. The quality of the output images will be evaluated through the Peak Signal to Noise Ratio (PSNR), Structural Similarity Index Method (SSIM), and the BRISQUE method assessments; subsequently, the output exhibiting the highest quality will be used to drive a crater detection model based on the YOLO v8 architecture to validate the results of crater detection against task-based metrics. In summary, the entire system constitutes an end-to-end imaging improvement and analytical pipeline and provides value for researchers studying lunar or planetary-based images.