Interaction-Guided Two-Branch Image Dehazing Network

TL;DR

This work tackles image dehazing by addressing the need to recover both local details and global scene consistency from hazy images modeled by $I = J(x)\,t(x) + A\,(1 - t(x))$ with $t(x) = e^{-{\beta} d(x)}$. It introduces an interaction-guided dual-branch network that couples a Transformer-based global perception with a CNN-based local perception, guided by a Channel-Pixel Attention (CPA) mechanism and strengthened by downsampling before the Transformer to reduce redundancy. The approach demonstrates competitive PSNR/SSIM/Entropy/LPIPS on RESIDE-6K and superior results on real datasets NH-HAZE and DENSE-HAZE, with ablation studies confirming the contribution of DownS, FA, and CPA. The work advances practical dehazing by delivering high-quality restoration with efficient computation and explicit cross-branch interaction.

Abstract

Image dehazing aims to restore clean images from hazy ones. Convolutional Neural Networks (CNNs) and Transformers have demonstrated exceptional performance in local and global feature extraction, respectively, and currently represent the two mainstream frameworks in image dehazing. In this paper, we propose a novel dual-branch image dehazing framework that guides CNN and Transformer components interactively. We reconsider the complementary characteristics of CNNs and Transformers by leveraging the differential relationships between global and local features for interactive guidance. This approach enables the capture of local feature positions through global attention maps, allowing the CNN to focus solely on feature information at effective positions. The single-branch Transformer design ensures the network's global information recovery capability. Extensive experiments demonstrate that our proposed method yields competitive qualitative and quantitative evaluation performance on both synthetic and real public datasets. Codes are available at https://github.com/Feecuin/Two-Branch-Dehazing

Figures (7)

• Figure 1: (a) hazy image, (b) and (c) represent the dehazing results of CNN method and the Transformer method, respectively, (d) Dehazing result of our method, and (e) groundtruth image.
• Figure 2: Our method utilizes Transformer and CNN branches, where the output features in the middle of each layer are utilized in a CPA generated attention map to guide the CNN. The CNN results are combined with the results of the Transformer branch to perform CNN upsampling to recover image details.
• Figure 3: Visual comparison of outdoor scenes of different dehazing methods on RESIDE-6K dataset. (Zooming in can obtain a clearer view)
• Figure 4: Visual comparison of indoor scenes of different dehazing methods on RESIDE-6K dataset. (Zooming in can obtain a clearer view)
• Figure 5: Visual comparison of different dehazing methods on NH-HAZE dataset.(Zooming in can obtain a clearer view)