UMD-NOIR: Unified Multiscale Diffusion Model for Navigation-Orientated Image Restoration

Abstract

Image restoration is essential for vision-based navigation, surveillance, and remote sensing, yet real-world images are often degraded by haze, fog, rain, snow, clouds, and underwater turbidity. Existing methods are typically tailored to single degradations or narrow domains, limiting their generalisation in complex environments with overlapping effects. We introduce UMD-NOIR, a unified multiscale diffusion model for navigation-orientated image restoration that integrates a Transformer-enhanced UNet into a denoising diffusion probabilistic model. The architecture incorporates multiscale channel and spatial attention together with a gated deconvolutional feed-forward network, enabling robust feature aggregation, spatially aware enhancement, detail recovery, and structural consistency. A hybrid loss combining Huber and perceptual terms further improves sharpness, colour accuracy, and perceptual fidelity. Quantitative evaluations against 13 state-of-the-art diffusion and transformer-based models show that UMD-NOIR achieves an average peak signal-to-noise ratio (PSNR) of 28.07 dB, structural similarity index measure (SSIM) of 0.888, and mean absolute error (MAE) of 0.032 across ground, aerial, and marine degradations. Generalisation is validated on five unseen datasets (DAWN, RESIDE, RICE, LSUI, and CDD-11), demonstrating adaptability to real-world and composite degradations. No-reference assessments further yield NIQE = 3.91 and PIQE = 20.85 on haze images, confirming perceptual realism. Downstream evaluations with YOLOv11 highlight improvements in classification, detection, and segmentation, while ablation studies verify the importance of multiscale processing, attention mechanisms, and hybrid loss in achieving consistent gains.