Prior-guided Hierarchical Harmonization Network for Efficient Image Dehazing

TL;DR

The paper tackles the challenging problem of single-image dehazing with a focus on practical efficiency. It introduces PGH2Net, a triple priors guided hierarchical network that integrates Bright Channel Prior, Dark Channel Prior, and Histogram Equalization through Prior Aggregation, Spatial/Channel Harmonization, a Sandwich bottleneck, and a Histogram Equation Guided Module. The approach delivers state-of-the-art or competitive PSNR/SSIM on RESIDE and real-world hazy datasets while reducing computational complexity, demonstrating strong restoration quality with lower FLOPs. This work advances dehazing by combining spatial guidance from BCP/DCP with distribution-guided HE and validates the effectiveness of hierarchical priors fusion for robust haze removal in real-world conditions.

Abstract

Image dehazing is a crucial task that involves the enhancement of degraded images to recover their sharpness and textures. While vision Transformers have exhibited impressive results in diverse dehazing tasks, their quadratic complexity and lack of dehazing priors pose significant drawbacks for real-world applications. In this paper, guided by triple priors, Bright Channel Prior (BCP), Dark Channel Prior (DCP), and Histogram Equalization (HE), we propose a \textit{P}rior-\textit{g}uided Hierarchical \textit{H}armonization Network (PGH$^2$Net) for image dehazing. PGH$^2$Net is built upon the UNet-like architecture with an efficient encoder and decoder, consisting of two module types: (1) Prior aggregation module that injects B/DCP and selects diverse contexts with gating attention. (2) Feature harmonization modules that subtract low-frequency components from spatial and channel aspects and learn more informative feature distributions to equalize the feature maps.

Figures (8)

• Figure 1: Reconstruction quality (PSNR) and computational complexity (FLOPs) on the SOTS-Indoor RESIDE dataset. The size of the dots indicates the model size.
• Figure 2: Visualization of the relationship between the spatial haze degradation(a)(d) and BCP(b) and DCP(c) in the deep feature domain. The error maps are differential values with reference, indicating haze distribution, shown in red boxes.
• Figure 3: Visualization of the relationship between the value distribution and feature channels. The second row shows histograms assigned to the first row of each image/feature. Similarities are calculated by Cosine similarity. The horizontal/x-axis is the normalized value from 0 to 1, and vertical/y-axis is the number of the value distribution of images/feature.
• Figure 4: PGH$^2$Net architecture. (a) The encoders and decoders with a stack of Prior Aggregation and Spatial/Channel Harmonization modules learn hierarchical features with diverse distributions. Then, the bottleneck with the (b) Sandwich Module (SM) and (c) Histogram Equation Guided Module (HEGM) transports equalized deep features to the decoders.
• Figure 5: Structure of the encoder and decoder blocks: Spatial Harmonization Module $\mathrm{SH}(\cdot)$, Prior Aggregation Module $\mathrm{PA}(\cdot)$, and Channel Harmonization Module $\mathrm{CH}(\cdot)$ are cascaded. $\mathrm{SH}(\cdot)$ and $\mathrm{PA}(\cdot)$ combine to aggregate spatial information.