FDDN: frequency-guided network for single image dehazing

Haze appearing in natural scene images generally contains nonhomogeneous characteristics such as filaments, masses, and mist. The high-frequency part of hazy images contains variable background textures and haze shapes, whereas regions with mostly uniform distribution are dominated by low-frequency information. Although existing methods based on convolutional neural networks have achieved remarkable progress in single image dehazing, the intrinsic hazy image patterns have been neglected in most models. We propose a frequency division dehazing network to leverage prior knowledge characterizing hazy images. The proposed network processes shallow feature maps through high-, medium-, and low-frequency branches. This separation facilitates a flexible architecture, whose branch handling lower-frequency components is less redundant given its relatively simpler background and haze shapes. Then, by integrating knowledge extracted from all the network branches using feature fusion, the proposed network fully exploits the variety of frequency characteristics in hazy images and achieves 39.51 PSNR and 0.9931 SSIM on the RESIDE dataset. Experiments on both synthetic and real hazy images demonstrate the superiority of the proposed network over several existing state-of-the-art methods, demonstrating the effectiveness of exploiting prior knowledge in hazy images.