Deep retinex-inspired enhancement of low-light lunar images using dark-bright image pairs

Lunar image analysis is very important to space research and automatic navigation but often face challenges while working on those images because of the extreme and non-terrestrial lighting conditions in space especially on moon. These conditions will result in producing low light images with deep shadows and low contrast, which are difficult for both human interpretation and computer vision. In our work, we introduced a Retinex inspired deep-learning architecture tailored for low light moon image enhancement, by using temporally aligned dark and bright image pairs. In our approach we used a dual network structure of Decom Net and Relight Net. Our results preserve structural accuracy while parallelly enhancing visual quality. Comprehensive evaluations involving key measures (PSNR: 29.00 dB, SSIM: 0.2851) during dataset creation and qualitative measures, confirm the performance superiority of our work and stand out of many solutions. Making it as appropriate tool for lunar image enhancement in space and operational environments.