Progressive Domain Translation Defogging Network for Real-World Fog Images

Abstract

Images captured in bad weather are affected by atmospheric scattering. To remove image degradation caused by scattering, many defogging methods have been proposed. However, due to the lack of consideration of high-frequency signal loss in real-world fog (RF) images, few existing methods are able to enhance RF images that have a large domain distance from mist images or synthetic fog images in the style feature space. Therefore, in this paper, a progressive domain translation defogging network is proposed to achieve the coordination of fog removal and target contour refinement for RF images. Firstly, a fog inversion module is trained with the newly built synthetic foggy images, and the high-frequency signal is truncated to shorten the distance between the training sample and RF images. The module realizes the restoration of fog density of RF images. Secondly, an image translation module is trained by an unsupervised loss to further eliminate fog degradation and enhance the style of defogged images toward clear images. The image style enhanced by the module includes the overall image style related to color, contrast, and the local image style related to target contours. Experimental results show that the proposed network can obtain high-quality RF images, which is beyond state-of-the-art methods in terms of visual quality indices and target detection indices. The code of the proposed network is found at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/yeyekurong/Guoqiang_PTD-Net/</uri> .