MiDUNet: Model-Inspired Deep Unfolding Network for Nonhomogeneous Image Dehazing

Le résumé fourni par la source

Optical remote sensing techniques, particularly those relying on visible optical sensors, are critical for Earth observations. However, atmospheric haze severely degrades image quality owing to light scattering effects. These effects often lead to clouds in the observed images, resulting in partial image degradation. To address the challenge of non-homogeneous image dehazing, we propose a model-inspired deep unfolding network (MiDUNet). The architecture of MiDUNet is strictly derived and inspired by a novel alternating minimization algorithm. During algorithm design, traditional image regularization terms are replaced by several nonlinear functions to align with neural network processing, increasing flexibility and applicability. Afterward, a Multi-scale Spatial frequency Residual Block (MSRB) is used to capture complex scene structures by fusing multi-scale spatial information and frequency information. With these spatial-frequency features and a Multi-path Feature Enhancement Block (MFEB), local and global information are obtained, which increases the robustness to non-homogeneous dehazing and image texture preservation. Numerical experiments with remote sensing images and natural images demonstrate the competitive advantages of MiDUNet in non-homogeneous dense haze scenarios.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.