Enhancing Underwater Image Using Degradation Adaptive Adversarial Network
Enhancing Underwater Image Using Degradation Adaptive Adversarial Network
复制标题
DOI：
10.1109/icip46576.2022.9897624
复制
发表时间：
2022-10
期刊：
影响因子：
--
通讯作者：
Lujun Zhai;Yonghui Wang;Suxia Cui;Yu Zhou
中科院分区：
文献类型：
--
作者：
Lujun Zhai;Yonghui Wang;Suxia Cui;Yu Zhou
水下成像由于光的折射和吸收，以及水中的悬浮颗粒，造成颜色失真和能见度差。现有的基于学习的水下图像复原方法高度依赖于成对的训练数据集。然而，在真实的水下场景中，失真/非失真图像对是不可用的，并且基于合成数据的训练模型在实际应用中的性能大大降低。提出了一种基于循环对抗生成网络(CycleGAN)和集成失真自适应(DA)模块的非配对水下图像复原方法。使用CycleGAN进行图像到图像的转换，DA模块可以灵活地从未配对的训练样本中学习各种退化信息。实验结果表明，该方法在颜色校正和视觉质量改善方面优于现有的水下图像复原算法。
Underwater imaging severely suffers from color distortion and poor visibility due to light refraction and absorption and suspended particles in the water. Existing learning-based underwater image restoration methods highly rely on paired training datasets. However, distorted/non-distorted image pairs in realistic underwater scenes are unavailable and synthetic data-based trained models greatly degrade the performance in real-world applications. In this paper, an unpaired underwater image restoration method is proposed using a cycle adversarial generative network (CycleGAN) with an integrated distortion adaptive (DA) module. CycleGAN is used to perform image-to-image translation and the DA module can learn various degradation information from the unpaired training samples with flexible adaption. Experimental results demonstrate the superiority of our proposed method against existing state-of-the-art algorithms in restoring underwater images in terms of color correction and visual quality improvement.