SIOD: Semantic Ingredient-Oriented Diffusion for Context-Aware Blind Image Restoration

الملخص

Blind Image Restoration (BIR) under unknown and mixed degradations remains a challenging problem due to the inherent trade-off between perceptual quality and semantic fidelity. Existing diffusion-based restoration methods typically rely on globally shared semantic priors and fixed diffusion schedules, which limit their ability to adapt restoration behavior according to regional semantic importance and uncertainty. Consequently, these methods often suffer from semantic inconsistency, hallucinated details, and performance degradation under severe or out-of-distribution (OOD) degradations.

To address these limitations, this paper proposes a Semantic Ingredient-Oriented Adaptive Diffusion (SIOD) framework for context-aware blind image restoration. The proposed framework introduces a Semantic Ingredient Decomposer that disentangles latent image representations into semantic-critical, structural, and texture-sensitive regions. A Region-wise Semantic Uncertainty Estimator is further employed to quantify restoration ambiguity and confidence across different image regions. Based on these uncertainty cues, an Adaptive Diffusion Scheduler dynamically allocates denoising strength, diffusion trajectories, and refinement depth according to the semantic importance and degradation severity of each region. To preserve semantic fidelity while maintaining perceptual quality, a Region-Adaptive Fidelity Calibration mechanism is incorporated to suppress hallucinated content and constrain unrealistic generation in sensitive regions. Finally, a Physics-Constrained Refinement Module integrates degradation-aware priors to enhance robustness and restoration stability under mixed and OOD degradations.

Extensive experiments on diverse blind image restoration benchmarks are expected to demonstrate superior semantic consistency, hallucination suppression, perceptual quality, and robustness compared with state-of-the-art diffusion-based restoration approaches.