CWE-Net: Cross weather expert network for adaptive image restoration

Image restoration is a fundamental task in computer vision, yet restoring outdoor scenes under multiple adverse weather conditions remains challenging due to heterogeneous and spatially coupled degradations that jointly affect visibility and scene consistency. Existing all-in-one restoration frameworks have achieved remarkable progress; however, unified architectures often entangle weather-specific cues, thereby hindering discriminative representation learning and interpretability. Motivated by the observation that different weather degradations often affect common scene structures (e.g., object boundaries and global illumination) while introducing distinct degradation cues such as contrast attenuation in haze, streak artifacts in rain, and particle occlusion in snow, we propose CWE-Net, an adaptive and unified framework that integrates weather-aware expert routing with cross-condition generalization, enabling the model to adapt across single, compound, and spatially mixed weather conditions. CWE-Net employs a lightweight multi-label router for explicit degradation perception, which dynamically activates three Base Weather Experts (BWEs) corresponding to haze, rain, and snow. Their outputs are refined through a Cross-Weather Expert (CWE) module that performs learned convolutional fusion and refinement guided by coordinate attention, enabling coherent reconstruction under compound and spatially inconsistent degradations. The network is optimized using a composite loss with lightweight adaptive weighting of structural, perceptual, and edge-level fidelity. Extensive evaluations on synthetic and real-world benchmarks demonstrate that CWE-Net achieves competitive performance, strong generalization, and clear interpretability, offering an effective and scalable framework for all-weather image restoration.