Guided Real Image Dehazing using YCbCr Color Space

TL;DR

This work tackles real-world image dehazing by addressing limitations of RGB-based methods in dense haze. It introduces SGDN, a dual-color-space framework that jointly processes RGB and YCbCr features, using a Bi-color Guidance Bridge (PIM and IAM) to transfer structural cues and a Color Enhancement Module to boost color perception, guided by a multi-scale loss. The Real-World Well-Aligned Haze (RW$^2$AH) dataset of 1,758 real paired images enables supervised learning and fair benchmarking against state-of-the-art methods. Empirical results show SGDN achieves superior dehazing performance on real-world datasets and provides notable improvements in downstream tasks, supported by extensive ablations and analyses. The work also contributes RW$^2$AH as a practical benchmark for evaluating real-world dehazing methods.

Abstract

Image dehazing, particularly with learning-based methods, has gained significant attention due to its importance in real-world applications. However, relying solely on the RGB color space often fall short, frequently leaving residual haze. This arises from two main issues: the difficulty in obtaining clear textural features from hazy RGB images and the complexity of acquiring real haze/clean image pairs outside controlled environments like smoke-filled scenes. To address these issues, we first propose a novel Structure Guided Dehazing Network (SGDN) that leverages the superior structural properties of YCbCr features over RGB. It comprises two key modules: Bi-Color Guidance Bridge (BGB) and Color Enhancement Module (CEM). BGB integrates a phase integration module and an interactive attention module, utilizing the rich texture features of the YCbCr space to guide the RGB space, thereby recovering clearer features in both frequency and spatial domains. To maintain tonal consistency, CEM further enhances the color perception of RGB features by aggregating YCbCr channel information. Furthermore, for effective supervised learning, we introduce a Real-World Well-Aligned Haze (RW$^2$AH) dataset, which includes a diverse range of scenes from various geographical regions and climate conditions. Experimental results demonstrate that our method surpasses existing state-of-the-art methods across multiple real-world smoke/haze datasets. Code and Dataset: \textcolor{blue}{\url{https://github.com/fiwy0527/AAAI25_SGDN.}}

Figures (13)

• Figure 1: Visual comparison of different color spaces: (a) RGB features degrade, blurring textures, while YCbCr is less affected by fog and shows clearer textures. (b) RGB models (e.g., MB-Taylor) leave residual haze, YCbCr models (e.g., AIPNet) distort colors. Our approach removes heavy fog while preserving color accuracy.
• Figure 2: Examples from the MRFID and BEDDE datasets show noticeable background differences between reference and haze images due to varying shooting angles and long time intervals. In contrast, our RW$^2$AH dataset achieves excellent alignment in physical space.
• Figure 3: The overall pipeline of our SGDN. It includes the proposed Bi-Color Guidance Bridge (BGB) and Color Enhancement Module (CEM). BGB promotes RGB features to produce clearer textures through YCbCr color space in both frequency and spatial domain, while CEM significantly enhances the visual contrast of the images.
• Figure 4: Geographic and haze distribution of our RW$^2$AH.
• Figure 5: Visual comparison results on the real-world smoke dataset. Zoom in for a better view.