Physical-Property Guided End-to-End Interactive Image Dehazing Network

Single image dehazing task predicts the latent haze-free images from hazy images corrupted by the dust or particles existed in atmosphere. Notwithstanding the great progress has been made by the end-to-end deep dehazing methods to recover the texture details, they usually cannot effectively preserve the real color of images, due to lack of constraint on color preservation. In contrast, atmospheric scattering model based dehazing methods obtain the restored images with relatively rich real color information due to unique physical property. In this paper, we propose to seamlessly integrate the properties of physics-based and end-to-end dehazing methods into a unified powerful model with sufficient interactions, and a novel Physical-property Guided End-to-End Interactive Image Dehazing Network (PID-Net) is presented. To make full use of the physical properties to extract the density information of haze maps for deep dehazing, we design a transmission map guided interactive attention (TMGIA) module to teach an end-to-end information interaction network via dual channel-wise and pixel-wise attention. This way can refine the intermediate features of end-to-end information interaction network, and do it a favor to obtain better detail recovery by sufficient interaction. A color-detail refinement sub-network further refines the dehazed images with abundant color and image details to obtain better visual effects. On several synthetic and real-world datasets, our method consistently outperforms other state-of-the-arts for detail recovery and color preservation.