Answers are grounded in this paper and cite their sources.
High-resolution image dehazing requires the recovery of broad scene structure together with fine directional details. We propose HLFDehazeNet, a wavelet-based network that regulates information exchange between these complementary representations. A discrete wavelet transform separates shallow features into one low-frequency band and three directional high-frequency bands, which are processed by parallel encoder–decoder branches with attentive state-space modules. Residual directional cross-frequency interaction (RDCI) uses a separate learned gate for each high-frequency band to modulate bidirectional exchange with the low-frequency branch before decoding. After inverse wavelet reconstruction, edge-conditioned residual compensation (ECRC) uses edge-related features to modulate an image-space residual correction. HLFDehazeNet obtains PSNR/SSIM values of 33.26 dB/0.9882 on 8KDehaze, 28.74 dB/0.9365 on 4KID, and 24.01 dB/0.7844 on O-HAZE. On RESIDE6K, the joint configuration improves PSNR by 2.56 dB over the same network with both modules disabled. The contribution of each module depends on the presence of the other, and internal ablations favor direction-specific gate predictors and residual modulation in the tested configurations. Without fine-tuning, HLFDehazeNet trained on RESIDE6K achieves PSNR/SSIM values of 28.97 dB/0.9773 on SOTS-Outdoor and 13.13 dB/0.5307 on NH-HAZE. Qualitative comparisons on the real-world subset of HSTS illustrate the restoration results on real hazy scenes.
No takes yet. Share an insight, caveat, or question.
Yan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: