DCE-Net: A Dual-Frequency Domain Knowledge-Guided Framework for Image Dehazing via Detail and Content Enhancements

Abstract

Existing image dehazing methods are largely constrained to spatial domain processing, failing to fully leverage the rich knowledge embedded in the frequency domain of clear images. Additionally, the traditional convolutional operations in network architectures limit their mapping capabilities to some extent. To address these issues, a novel image dehazing network, termed the Detail and Content Enhancement Network (DCE-Net), is proposed. DCE-Net redefines dehazing task from a frequency-domain perspective, incorporating differential convolution and attention mechanisms to design the High-Frequency Detail Enhancement Module (HDEM) and the Low-Frequency Content Enhancement Module (LCEM). Furthermore, a Dual-Frequency Domain Knowledge-Guided Strategy (DDKS) is introduced during the training phase to exploit the abundant frequency-domain priors inherent in clear images. Experimental results demonstrate that the DCE-Net achieves outstanding performance on both synthetic benchmark datasets and real-world hazy scenes. DCE-Net not only significantly restores image clarity and contrast but also effectively preserves details and content features.