Rich feature distillation with feature affinity module for efficient image dehazing

Single-image haze removal is a long-standing hurdle for computer vision applications. Several works have been focused on transferring advances from image classification, detection, and segmentation to the niche of dehazing, primarily focusing contrastive learning knowledge distillation. However, these approaches prove computationally expensive, raising concern regarding their applicability on-the-edge use-cases. This work introduces simple, lightweight, efficient framework single-image removal, exploiting rich “dark-knowledge” information lightweight pre-trained super-resolution model via notion heterogeneous We designed feature affinity module maximize flow semantics teacher student dehazing network. In order evaluate efficacy our proposed framework, its performance as plug-and-play setup baseline examined. Our experiments are carried out RESIDE-Standard dataset demonstrate robustness synthetic real-world domains. The extensive qualitative quantitative results provided establish effectiveness achieving gains up 15% (PSNR) while reducing size by ?20 times.