S3Conv: Element-wise Adaptive Soft Structural Sharing for Multi-Weather Image Restoration

To ensure the dependability of outdoor computer vision applications, it is crucial to recover clear visual data from weather-corrupted images. Multi-weather image degradations are characterized by a dual nature: the presence of shared structural priors (e.g., textures and contours) amidst distinct, particle-dependent degradation patterns. Capturing the correlation of the former while addressing the specificity of the latter is essential for establishing a robust and unified restoration framework. Although a unified restoration paradigm is essential to capture these dual factors, current models predominantly employ coarse-grained hard-sharing schemes. Such architectures are inherently suboptimal, as they fail to effectively disentangle commonalities from specificities, thereby inducing performance-degrading negative transfer. To this end, we propose a fine-grained parameter sharing strategy based on Structured SoftSharing Convolution, termed S3Conv. S3Conv employs shared kernels to capture universal texture perturbations and specific kernels to model distinct physical patterns, and fuses them at the element level using an input-adaptive gating tensor. The gate is jointly driven by channel-wise covariance, which encodes statistical structure, and spatial gradient sparsity, which quantifies texture complexity. Importantly, the resulting complexity estimate produces a texture-aware Gaussian kernel that dynamically modulates the effective receptive field, allowing the network to adapt between modeling localized high-frequency noise and global degradation distributions. Benchmarking on diverse public datasets shows that the proposed architecture achieves strong restoration performance and demonstrates the potential of fine-grained soft sharing for unified multi-weather restoration.