HFIP: High-Frequency Information Prompt Network for Single Image Dehazing

Jufeng Li, Yuxin Feng, Mouguang Lin, Zhuo Su

Abstract: Current dehazing methods perform well on synthetic haze datasets but often struggle with complex real-world scenarios. To enhance the model's generalization ability, we incorporate Fourier transform and Gaussian filter to capture frequency domain information. Based on this, we propose a high-frequency information prompt network to more effectively locate haze and restore texture details. Additionally, to address the issue of overly dark dehazing results commonly observed in networks trained on synthetic datasets, we introduce a brightness loss function based on threshold division, which effectively enhances the brightness of the output images. Through extensive experiments, our method demonstrates not only excellent dehazing performance in challenging and complex scenarios such as daytime fog and sandstorm, but also significant potential value for downstream applications.