DehazeDCT: Towards Effective Non-Homogeneous Dehazing via Deformable Convolutional Transformer

TL;DR

This work targets non-homogeneous hazy image restoration at high resolutions by introducing DehazeDCT, a two-stage architecture that combines a DCNv4-based Transformer-like dehazing module with a light Retinex-inspired refinement. The Dehazing module employs DCNFormer blocks to capture long-range dependencies and adaptive spatial aggregation without self-attention's computational burden, aided by a frequency-aware branch. A separate Refinement module refines color and textures via a lightweight Retinex-inspired transformer, improving color fidelity and detail without adversarial training. Empirical results on NH-HAZE, NH-HAZE2, HD-NH-HAZE, and DNH-HAZE2 demonstrate strong PSNR/SSIM gains and competitive performance in the NTIRE 2024 Dense and Non-Homogeneous Dehazing Challenge, highlighting the approach's practicality for high-resolution non-homogeneous dehazing.

Abstract

Image dehazing, a pivotal task in low-level vision, aims to restore the visibility and detail from hazy images. Many deep learning methods with powerful representation learning capability demonstrate advanced performance on non-homogeneous dehazing, however, these methods usually struggle with processing high-resolution images (e.g., $4000 \times 6000$) due to their heavy computational demands. To address these challenges, we introduce an innovative non-homogeneous Dehazing method via Deformable Convolutional Transformer-like architecture (DehazeDCT). Specifically, we first design a transformer-like network based on deformable convolution v4, which offers long-range dependency and adaptive spatial aggregation capabilities and demonstrates faster convergence and forward speed. Furthermore, we leverage a lightweight Retinex-inspired transformer to achieve color correction and structure refinement. Extensive experiment results and highly competitive performance of our method in NTIRE 2024 Dense and Non-Homogeneous Dehazing Challenge, ranking second among all 16 submissions, demonstrate the superior capability of our proposed method. The code is available: https://github.com/movingforward100/Dehazing_R.

Figures (6)

• Figure 1: Test Result of our method on NTIRE 2024 Dense and Non-Homogeneous Dehazing Challenge NTIRE_Dehazing_2024. Our DehazeDCT achieves the second best performance among 16 solutions and is capable to generate visually compelling outputs with vivid color and enhanced structure details.
• Figure 2: The overall architecture of our proposed model. In the Dehazing module, we introduce a transformer-like dehazing branch based on deformable convolution (DCNv4 xiong2024efficient). In each DCNFormer block, DCNv4 is utilized to calculate the offset ($\Delta p$) and modulation scalar ($\mathbf{m}$). Besides, the frequency-aware branch proposed in DWT-FFC_2023_CVPRW is also adopted as an auxiliary branch. In the Refinement module, we leverage a lightweight retinex-inspired transformer network to further reduce the color deviation and enhance texture details.
• Figure 3: Visual comparisons on NH-HAZE NH-Haze_2020 dataset. Compared to other models, our method exhibits higher color fidelity and effective dehazing, yielding compelling results.
• Figure 4: Visual experiment results on NH-HAZE NTIRE_Dehazing_2021 dataset. Obviously, our method demonstrates superior performance on color preservation and detail maintaining, further enhancing the overall quality of the output.
• Figure 5: Visual Comparisons on HD-NH-HAZE dataset NTIRE_Dehazing_2023. Our method exhibits superior haze removal, evidenced by more vivid colors and clearer details, especially in the foliage and background structures.