M-CBN: Manifold constrained joint image dehazing and super-resolution based on chord boosting network

Abstract

This paper proposes a Manifold Constrained Chord Boosting Network (M-CBN), which incorporates the super-resolution principle to achieve image dehazing. M-CBN is a task-specific image restoration network that explicitly learns the mapping from Low-resolution (LR) hazy images to High-resolution (HR) haze-free images. Hence, we design a preliminary image degradation to imitate super-resolution training on hazy images. In M-CBN, a plug-and-play Cross-linked Dual Projection Module (CDPM) for skip connections is developed. In CDPM, back-projections for HR encoder features and LR decoder features are cross-linked for better recovery of spatial information, and a cross-resolution spatial attention is designed to enhance fusion features. Then, to boost the generation of image details and textures, we propose a Chord Residual Module (CRM), which can separately process High-frequency (HF) and Low-frequency (LF) features by progressive inner-frequency updating and dense inter-frequency cross-collaboration to enhance decoding features. Finally, a manifold constraint dual discriminator is established. The static discriminator explicitly constrains dehazed images in the expected manifold to unify the joint learning of image dehazing and super-resolution. And the dynamic discriminator implicitly optimizes the network by adversarial training. Extensive experiments on general, dense and non-homogeneous haze datasets and cross-domain dehazing tasks show the proposed M-CBN presents high-quality dehazed results with natural colors and clear details.