Proceedings of the Annual Meeting of the Cognitive Science Society

Modeling perceptual restoration under severe degradation remains a core challenge in vision science. Inspired by the hierarchical organization of human episodic memory, we propose a cognitively grounded two-stage diffusion framework for single image dehazing. Analogous to neocortical–hippocampal division of labor, the first stage encodes a coarse, stable global structure, while the second stage progressively restores fine-grained details conditioned on this intermediate representation. This staged reconstruction mirrors memory consolidation, where global context is stabilized before perceptual detail is refined, enabling a more efficient and cognitively plausible coarse-to-fine inference process. By decomposing restoration into functionally distinct stages, our framework avoids inefficient multi-scale optimization and promotes stable, coherent convergence through target-conditioned diffusion dynamics. Extensive experiments demonstrate that the proposed model achieves state-of-the-art performance, exhibiting strong robustness, interpretability, and practical applicability in complex real-world dehazing scenarios.