DDTNet: Degradation Disentanglement and Transfer Network for Domain-Adaptive All-in-One Image De-weathering

Kuan-Hung Lin, Fu-Jen Tsai, Yan-Tsung Peng, Min-Hung Chen, Chia-Wen Lin, Yen-Yu Lin

Keywords: Domain Adaptation, Image Restoration

Abstract: All-in-one adverse weather image restoration aims to remove multiple degradations, such as rain, haze, and snow, using a single model. Despite their broad applicability, existing methods typically compromise performance, delivering balanced rather than optimal results for individual degradations due to their multi-task nature. Moreover, they often suffer from a significant performance drop when a domain gap exists between training and testing data. To address these challenges, we propose the Degradation Disentanglement and Transfer Network (DDTNet), which carries out domain adaptation for all-in-one models. Since paired degraded-clean images are unavailable at inference, DDTNet disentangles and transfers degradation patterns from target-domain degraded images to source-domain clean images, generating domain-adaptive pairs for fine-tuning and improving target-specific restoration. The core of DDTNet is the Degradation Disentanglement Module (DDM), which consists of Degradation Coupled Attention (DCA) to capture both general and weather-specific features, enabling effective disentanglement and transfer of degradation patterns. Experimental results demonstrate that DDTNet significantly improves existing all-in-one models across real-world deraining, desnowing, and dehazing datasets.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 2760