Single UHD image dehazing via Interpretable Pyramid Network

Abstract

Currently, most dehazing-based deep approaches are developed as an end-to-end manner to reconstruct a degraded image in an unintelligible fashion. For these dehazing models, the absence of ‘Credible Modeling’ to guide the network design is a barrier to be applied commercially in the open-world. To address this problem, we introduce the Taylor approximation principle as the soul and materialize this principle with the help of the Laplacian Pyramid. Specifically, we assume that the N paths of Laplacian pyramid model correspond to the N terms (sub-functions) in Taylor’s theorem. We attempt to use bottom paths and top paths to reconstruct the low-/high-frequency information of the clear image, respectively. Further, we develop a T-Unet module that focuses on regularizing the feature maps generated at bottom paths and design an attention sharing weight K to help approximate the Taylor high-order terms. Extensive experimental results demonstrate that our approach can run a 4K image on a single GPU with 24G RAM in real-time (80fps) and have unparalleled interpretability. The url of our code at https://github.com/zzr-idam/Interpretable-Pyramid-Network.