DDR-Net: Haze-Aware Dual-Domain Refinement for Single-Image Dehazing

Jul 16, 2026

Upload by Quinn

Single-image dehazing targets the recovery of clear scenes from haze-degraded images, but remains difficult because atmospheric scattering and diverse real-world haze distributions hinder accurate restoration. While end-to-end networks show progress, bottleneck-stage feature refinement and local structural representation in encoder–decoder designs limit performance. DDR-Net addresses this with three modules: HPE for multi-scale haze-aware priors, DE Blocks with gradient-aware convolutions for enhanced edges and textures, and SFBR that jointly refines spatial and frequency bottleneck features. Experiments on real-world benchmarks show superior haze removal and competitive synthetic results.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages15

Reading Time38 min

Introduction

Background and motivation

Related work

Proposed method and contributions

arXiv :2607 . 11071v1 [ cs .CV] 13 Jul 2026 DDR-Net: Haze-Aware Dual-Domain Refinement for Single-Image Dehazing Xinye Zheng, Ye Yu⋆ , Qiang Lu, Jinsheng Luo, Yiran Cui, and Yongbin Cheng School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China [yuye@hfut.edu.cn](yuye@hfut.edu.cn) Abstract. Single-image dehazing aims to recover clear scenes from hazedegraded images. It remains challenging due to the atmospheric scattering and the complexity of real-world haze distributions. Although recent end-to-end networks have achieved promising performance, two issues still limit their effectiveness: insufficient feature refinement at the bottleneck stage and weak local structural representation in encoder-decoder architectures. Thus, we propose a Haze-Aware Dual-Domain Refinement Network (DDR-Net) for single-image dehazing. Our method is built upon three modules: Haze Prior Extractor (HPE) provides multi-scale hazeaware priors by operating directly on downsampled hazy images; DetailEnhanced Blocks (DE Blocks) serve as the core feature extraction units, capturing multi-scale structural information and enhancing edge and texture recovery via gradient-aware convolutions; and Spatial-Frequency Bottleneck Refinement (SFBR) at the bottleneck jointly exploits spatial and frequency information to refine bottleneck features. DDR-Net achieves more effective feature representation and reconstruction for haze removal. Extensive experiments on real-world benchmarks demonstrate that our method outperforms existing dehazing approaches. It achieves competitive performance on synthetic datasets. Keywords: Single Image Dehazing · Image Restoration · Low-level Vision 1 Introduction Single-image dehazing aims to recover a clear scene from a haze-degraded image. It serves as a critical preprocessing step for a variety of practical applications, including autonomous driving and remote sensing. In hazy environments, particles such as aerosols and water droplets scatter and absorb incident light, which leads to low contrast and color distortion. Meanwhile, these degradations introduce a negative impact on visual quality and downstream vision tasks such as object detection, semantic segmentation, and scene understanding. Thus, developing effective image dehazing methods is important for both academia and industry. ⋆ Corresponding author 2 X. Zheng et al. | DD | R-Net (Ours | ) | MB-Taylo | rFormer-LV2 | | | --- | --- | --- | --- | --- | --- | | | | | MB-Tay | lorFormer-LV1 | | | | | ConvI<br> | R-B | FSNet<br> | ConvIR-L | | | ConvIR-S | | PMNet | | | | | | | | | | | | | | | | | | | | | | | | 40 60 80 100 120 140 MACs (G) 35.5 35.0 34.5 34.0 33.5 33.0 32.5 32.0 | DDR-Net | (Ours) MB-Taylo | rFormer-LV2 | | | --- | --- | --- | --- | | | MB-Taylo | rFormer-LV1 | ConvIR-L | | | Conv<br> | IR-B FSNet<br> | | | | ConvIR-S | | | | | | | PMNet | | | | | | | | | | | 0 5 10 15 20 Params (M) Fig. 1. Comparison of representative SOTA dehazing methods on the Haze4K dataset. Left: MACs (G) vs. PSNR (dB); Right: Params (M) vs. PSNR (dB) . Our DDR-Net achieves competitive dehazing performance with lower computational cost and fewer parameters than recent SOTA methods. Early dehazing approaches relied on handcrafted priors, such as Dark Channel Prior (DCP) [11] and Color Attenuation Prior (CAP) [40] . Although these methods can achieve satisfactory results under certain assumptions, their performance often degrades in complex real-world scenarios due to limited generalization ability and sensitivity to parameter settings. With the rapid development of deep learning, CNN-based dehazing methods have significantly advanced the field [4,27,25,15,10,5] . DehazeNet [4] and MSCNN [27] employed convolutional networks to estimate transmission maps, while AOD-Net [15] reformulated the atmospheric scattering model for end-to-end image restoratio

FAQs

What problem does DDR-Net address?

DDR-Net addresses single-image dehazing, aiming to recover a clear scene from haze-degraded images where atmospheric scattering causes low contrast and color distortion.

Why are current end-to-end dehazing networks limited?

The document states two key limitations: insufficient feature refinement at the bottleneck stage and weak local structural representation in encoder–decoder architectures.

What modules compose DDR-Net?

DDR-Net consists of Haze Prior Extractor (HPE), Detail-Enhanced (DE) Blocks, and Spatial-Frequency Bottleneck Refinement (SFBR). HPE extracts multi-scale haze-aware priors, DE Blocks enhance local structure via gradient-aware convolutions, and SFBR jointly refines features using spatial and frequency information.

Jul 16, 2026

Upload by Quinn