Unsupervised Image Dedusting via a Cycle-Consistent Generative Adversarial Network

Abstract

In sand–dust weather, the quality of the image is seriously degraded, which affects the ability of advanced applications to image using remote sensing. To improve the image quality and enhance the performance of image dedusting, we propose an end-to-end cyclic generative adversarial network (D-CycleGAN) for image dedusting, which does not require pairs of sand–dust images and corresponding ground truth images for training. In other words, we train the network in an unpaired way. Specifically, we designed a jointly optimized guided module (JOGM), comprised of the sandy guided synthesis module (SGSM) and the clean guided synthesis module (CGSM), which aim to jointly guide the generator through corresponding discriminator adversarials to reduce the color distortion and artifacts. JOGM can significantly improve image quality. We propose a network hidden layer adversarial branch to perform adversarials from inside the network, which better supervises the hidden layer to further improve the quality of the generated images. In addition, we improved the original CycleGAN loss function and propose a dual-scale semantic perception loss in feature space and a color identity-preserving loss in pixel space to constrain the network. Extensive experiments demonstrate that our proposed network model effectively removes sand dust, has better clarity and image quality, and outperforms the state-of-the-art techniques. In addition, the proposed method can help the target detection algorithm to improve its detection accuracy and capability, and our method generalizes well to the enhancement of underwater images and hazy images.