Work

CT-Weather: A Unified CNN–Transformer Hybrid Architecture for Multi-Weather Image Restoration with Explicit Degradation Cues

Year: 2026

Type: article

Abstract: Restoring images captured under adverse weather conditions is essential for robust computer vision applications. However, most existing methods struggle to handle diverse and challenging weather degradation with a single model. Although some approaches leverage prompt learning for weather adaptability, prompts implicitly learned from restoration objectives may provide limited degradation guidance. Meanwhile, insufficient extraction, propagation, and reuse of shallow edges and textures often lead to over-smoothed outputs and artifacts. To address these issues, we propose CT-Weather, an effective unified encoder–decoder architecture explicitly injected with rich degradation cues for multi-weather restoration. The architecture consists of three key components: (i) Feature Aggregation Attention Block (FAAB), which effectively extracts fine-grained details and facilitates efficient feature propagation and reuse; (ii) Degradation Pattern Attention Module (DPAM), which fully exploits multi-level features ranging from low-level textures to high-level semantics to adaptively model and capture differentiated spatial degradation patterns; and (iii) Momentum-Proxy Adaptation (MPA), a novel metric learning strategy that trains an auxiliary network to learn highly discriminative and stable weather-degradation embeddings. Extensive experiments on multi-weather benchmarks show that CT-Weather achieves state-of-the-art performance, surpassing 33.6 dB in average peak signal-to-noise ratio (PSNR) and 0.952 in average structural similarity index measure (SSIM). Furthermore, evaluations on real-world images validate its strong generalization capability. Notably, even in comparison with Nano Banana Pro (Gemini 3 Pro Image), a large-scale pre-trained image generation and editing model, CT-Weather demonstrates superior restoration quality and structural fidelity.

Source: IEEE Transactions on Circuits and Systems for Video Technology

DOI: 10.1109/tcsvt.2026.3748246

Authors: Jianghui Ma, Xiyuan Chen, Kaiwei Sang, Weiming Jing, Xuyang Jiang, Shaojun Feng

Institution: Southeast University

Language: English

Topic: Image Enhancement Techniques

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

SDG: Climate action

Open Access status: hybrid

Estimated APC paid: $2,645