D 2 W-net: dynamic wavelet prior and dual-domain learning dehazing network for enhanced object detection in traffic scenes

Abstract Haze-induced image degradation poses a significant challenge to high-level vision tasks, particularly object detection in intelligent transportation systems. While recent dehazing methods have significantly improved conventional image quality metrics, they often fall short in enhancing downstream detection performance. Moreover, these methods lack comprehensive modeling of haze degradation in the frequency domain. To address these issues, this study investigates haze degradation from a wavelet-domain perspective and observes its scale-dependent characteristics. Specifically, haze-related information is predominantly concentrated in the low-frequency LL subbands, where different hierarchical levels capture degradation priors ranging from fine-grained light haze to coarse-grained dense haze. Building upon these insights, we propose D 2 W-Net (Dynamic Wavelet Prior and Dual-Domain Learning Dehazing Network). The model employs Dynamic Wavelet Depthwise Separable Convolution (DyWConv) to enable content-adaptive feature encoding. It effectively unites the complementary strengths of wavelet and Fourier transforms under the MetaFormer architecture paradigm. Extensive experiments demonstrate that D 2 W-Net achieves leading dehazing performance on synthetic datasets, attaining PSNR scores of 42.79 dB on SOTS-Indoor and 38.92 dB on SOTS-Outdoor, as well as superior results on real-world hazy images. More importantly, on the Foggy Cityscapes dense haze subset ( β =0.02), YOLOv8l equipped with our D 2 W-Net achieves a 7.2 percentage point improvement in mAP@0.5 compared to the original hazy images. This result outperforms existing approaches and highlights its practical utility for robust perception in hazy traffic environments.