Dense hazy image dehazing network with progressive learning paradigm and frequency decoupling enhancement

Dense hazy image dehazing is a challenging task. When processing dense haze images, the multi-layer encoding compression of deep model often leads to the loss of originally high-frequency features. Under traditional supervised learning paradigms, it is difficult to obtain a clear image from a dense hazy one, and the convergence of model training cannot be guaranteed. To address these issues, we propose a novel U-Net-based model with frequency decoupling enhancement (FDE) to dehaze dense hazy images. The FDE decouples the multi-level frequency features of dense hazy images, preserving an image’s primary information and enhancing high-frequency details. The spatial-frequency interaction (SFI) module fuses high-level frequency features with spatial features, effectively making them complement each other. Meanwhile, the noise suppressor (NS) is designed to reduce the high frequency noise derived by FDE. Our progressive learning paradigm draws inspiration from transfer learning, where pretraining is conducted on a simplified version of the complex target task. This approach involves training a generative model to convert dense hazy images into light hazy images, followed by fine-tuning the model’s parameters to adapt to the more complex dense haze removal task. This strategy prevents training collapse during dense haze removal. Experimental results demonstrate that the proposed method achieves favorable subjective and objective performance across various dense hazy image dehazing datasets. The code for this work is available at https://github.com/Paris0703/progressive_dehazing.