Adapt CLIP as Aggregation Instructor for Image Dehazing

Most dehazing methods suffer from limited receptive field and do not explore the rich semantic prior encapsulated in vision-language models, which have proven effective in downstream tasks. In this paper, we introduce CLIPHaze, a pioneering hybrid framework that synergizes the efficient global modeling of Mamba with the prior knowledge and zero-shot capabilities of CLIP to address both issues simultaneously. Specifically, our method employs parallel state space model and window-based self-attention to obtain global contextual dependency and local fine-grained perception, respectively. To seamlessly aggregate information from both paths, we introduce CLIP-instructed Aggregation Module (CAM). For non-homogeneous and homogeneous haze, CAM leverages zero-shot estimated haze density map and high-quality image embedding without degradation information to explicitly and implicitly determine the optimal neural operation range for each pixel, thereby adaptively fusing two paths with different receptive fields. Extensive experiments on various benchmarks demonstrate that CLIPHaze achieves state-of-the-art (SOTA) performance, particularly in non-homogeneous haze. Code will be publicly after acceptance.