DFR-Net: Density Feature Refinement Network for Image Dehazing Utilizing Haze Density Difference

TL;DR

This work tackles image dehazing under varying haze densities by introducing DFR-Net, a density-aware network that learns from density differences between a hazy image and a lower-density proposal image. It uses a two-branch architecture (Global Branch and Local Branch) with a Proposal Image Generator, a Global Density Feature Refinement module, and a Local Density refinement pipeline including S&M, DAFF, and IDRF to extract and refine density features globally and locally. The model optimizes with reconstruction, perceptual, and density-focused losses, and experiments show state-of-the-art results on RESIDE-outdoor, Haze4K, Dense-Haze, and NH-HAZE, with strong ablation support for the design choices. The approach provides a robust, interpretable density-aware dehazing framework that does not require T-map annotations and demonstrates improved detail preservation across diverse hazy scenes, offering practical impact for real-world dehazing tasks.

Abstract

In image dehazing task, haze density is a key feature and affects the performance of dehazing methods. However, some of the existing methods lack a comparative image to measure densities, and others create intermediate results but lack the exploitation of their density differences, which can facilitate perception of density. To address these deficiencies, we propose a density-aware dehazing method named Density Feature Refinement Network (DFR-Net) that extracts haze density features from density differences and leverages density differences to refine density features. In DFR-Net, we first generate a proposal image that has lower overall density than the hazy input, bringing in global density differences. Additionally, the dehazing residual of the proposal image reflects the level of dehazing performance and provides local density differences that indicate localized hard dehazing or high density areas. Subsequently, we introduce a Global Branch (GB) and a Local Branch (LB) to achieve density-awareness. In GB, we use Siamese networks for feature extraction of hazy inputs and proposal images, and we propose a Global Density Feature Refinement (GDFR) module that can refine features by pushing features with different global densities further away. In LB, we explore local density features from the dehazing residuals between hazy inputs and proposal images and introduce an Intermediate Dehazing Residual Feedforward (IDRF) module to update local features and pull them closer to clear image features. Sufficient experiments demonstrate that the proposed method achieves results beyond the state-of-the-art methods on various datasets.

Figures (11)

• Figure 1: The main idea of our DFR-Net. By utilizing the density difference between the input (I) and the generated proposal image (P), global and local density features are refined in different ways: features of different global densities are pushed farther apart and local density features are pulled in towards clear image features.
• Figure 2: The pipeline of DFR-Net. DFR-Net first generates a proposal image ($\textbf{P}$) by PIG, and $\textbf{P}$ is input to the subsequent two-branch network together with the hazy input ($\textbf{I}$). Each branch predicts a pseudo-clear image and we perform an adaptive fusion to obtain the final result. Note that the global density features in Global Branch are fed into Local Branch by cross-branch connections.
• Figure 3: The illustration of GB, global block and GDFR module. Note the global density features are fed into LB by cross-branch connections.
• Figure 4: The illustration of the structure of LB (top) and DAFF (a) , S&M (b), CSDA (c) modules. The locations of the IDRF usage are indicated by dashed lines representing plug-and-play availability.
• Figure 5: The illustration of IDRF module. The projected feature $F^{'}_L$ will be used to update local features in corresponding S&M or DAFF module.