FDG-Diff: Frequency-Domain-Guided Diffusion Framework for Compressed Hazy Image Restoration

TL;DR

This work tackles the practical challenge of restoring JPEG-compressed hazy images by proposing FDG-Diff, a frequency-domain-guided restoration framework that integrates spectrum decomposition in the DCT domain with a compression-aware diffusion process. A High-Frequency Compensation Module (HFCM) uses Haar-wavelet features and cross-attention with a predicted compression spectrum to recover lost high-frequency details, while the Degradation-Aware Denoising Timestep Predictor (DADTP) enables region-specific timesteps via patch-level transmission maps. The method employs a Spectrum Decomposition Network to separate compression effects and corrected hazy content, guiding the diffusion model through cross-layer information. Experimental results on four JPEG hazy datasets show consistent improvements over state-of-the-art methods, with notable gains in dense and uneven haze scenarios, highlighting the approach's practical impact for real-world compressed hazy image restoration.

Abstract

In this study, we reveal that the interaction between haze degradation and JPEG compression introduces complex joint loss effects, which significantly complicate image restoration. Existing dehazing models often neglect compression effects, which limits their effectiveness in practical applications. To address these challenges, we introduce three key contributions. First, we design FDG-Diff, a novel frequency-domain-guided dehazing framework that improves JPEG image restoration by leveraging frequency-domain information. Second, we introduce the High-Frequency Compensation Module (HFCM), which enhances spatial-domain detail restoration by incorporating frequency-domain augmentation techniques into a diffusion-based restoration framework. Lastly, the introduction of the Degradation-Aware Denoising Timestep Predictor (DADTP) module further enhances restoration quality by enabling adaptive region-specific restoration, effectively addressing regional degradation inconsistencies in compressed hazy images. Experimental results across multiple compressed dehazing datasets demonstrate that our method consistently outperforms the latest state-of-the-art approaches. Code be available at https://github.com/SYSUzrc/FDG-Diff.

Figures (6)

• Figure 1: Improvement of FDG-Diff over the SOTA approaches on I-Haze ihaze dataset. The circle size represents the corresponding SSIM psnrssim values. Higher PSNR and SSIM indicate better performance, while lower FID values are preferred.
• Figure 2: The track of the JPEG process. The colors in the loss map reflect the magnitude of the loss, as the absolute loss is computed separately for each of the three channels. Though compressed at a quite high QF of 80, the hazy images still suffers severe information loss in hazy regions.
• Figure 3: (a) The pipeline of FDG-Diff, comprising a spectrum decomposition network and a compression-aware frequency compensation DDPM. The spectral decomposition network separates compression effects in the frequency domain and generates a corrected hazy image, which serves as the conditional guidance for the DDPM. Additionally, the Degradation-Aware Denoising Timestep Predictor (DADTP) adjusts the denoising timesteps for each patch to enable region-specific restoration. (b) The detailed block design of the HFCM. Input features are enhanced in the mid-to-high frequency band using the wavelet transform and processed through cross-attention cross_attention with the compression effect spectrum to achieve precise compensation for the signal loss caused by compression.
• Figure 4: The structure of DADTAP. The DADTP employs a spatial-channel dual-attention mechanism to effectively fuse features, adaptively adjusting the timestep offset for each patch.
• Figure 5: Visual comparison on I-Haze-JPEG ihaze, O-Haze-JPEG ohaze, Dense-Haze-JPEG densehaze, and NH-Haze-JPEG nhaze, with all inputs compressed at QF 80. *Denotes the adoption of the superior cascading strategy. Our method yields fewer artifacts, more realistic details and better color consistency than others.