Adaptive Dual Fusion Network for All-in-One Image Restoration

Addressing degradations in a unified network is the goal of all-in-one image restoration. Due to the diversity of corruptions such as rain, haze, and noise, this is a major challenge. Although existing methods have made significant progress, they primarily operate in spatial domain, ignoring the unique characteristics that different degradations exhibit in frequency domain. To this end, we propose an adaptive dual fusion network (ADFNet), which aims to synergistically utilize features from both spatial and frequency domains. Adaptive dual fusion module (ADFM) is the central component of ADFNet. ADFM leverages spatial and frequency features through spatial branch and frequency branch, dynamically fusing them through adaptive gating. Extensive experiments demonstrate that ADFNet achieves state-of-the-art performance in image restoration.