Region of Interest-focused Dynamic Enhancement (RoIDE) for Satellite Images

A major challenge in satellite imaging is the wide-ranging brightness levels within a single frame, which can lead to inadequate exposure in shaded regions and overexposure in bright areas. This research proposes a two-stage deep learning architecture, region of interest-focused dynamic enhancement (RoIDE), for multiexposure generation (MEG) and fusion. The first stage generates multiple exposures from a single standard dynamic range (SDR) image, capturing details across different brightness levels. The second stage fuses these images to create a high-contrast, comprehensive image, preserving details in both low and high-exposure regions. This approach allows for focused processing, enhancing regions of interest (RoI) without causing overexposure. Experimental results show significant improvements in the dynamic range of satellite images, particularly in RoIs. Enhanced visibility and detail preservation in both dark and bright areas are achieved, demonstrating the effectiveness of the RoIDE architecture. Nonreference metrics, including perception-based image quality evaluator (PIQE) (34.92), and blind tone-mapped quality index (BTMQI) (0.54), confirm the model’s superior perceptual, spatial, naturalness, and tone-mapped quality. Additionally, image quality assessments using peak signal-to-noise ratio (PSNR) (30.43), learned perceptual image patch similarity (LPIPS) (0.168), and VDP-3 (9.35) show significant improvements over other state-of-the-art models. These advancements are crucial for accurate data analysis in applications relying on satellite imagery.