基于域自适应的红外图像去噪算法

在红外图像去噪任务中，由于真实的红外噪声图像难以大量获取，而使深度学习算法高度依赖于人工合成噪声，无法很好地去除真实的红外噪声。本文提出一种基于域自适应的红外图像去噪算法，包括一个图像转换模块和两个图像去噪模块。首先利用图像转换模块将合成红外噪声图像和真实红外噪声图像相互转换，然后将转换后的图像和原图像作为去噪模块的训练数据，采用一致性损失函数使两个图像去噪模块产生一致的结果，最后将训练后的去噪网络框架用于红外图像去噪任务。实验表明，本文提出的算法与BM3D、DnCNN和ADNet算法相比在合成红外噪声数据集上有更高的指标数值和更好的视觉效果，在真实红外噪声数据集上有同样优秀的去噪效果。证明了该算法具有良好的泛化能力，能够在真实噪声下恢复清晰的红外图像。

Infrared image denoising algorithm based on domain adaptation

In the infrared image denoising task,the deep learning algorithms are highly dependent on synthetic noise and cannot remove the real infrared noise well since the real infrared noise images are difficult to obtain in large batches.In this paper,a network framework for infrared image denoising based on domain adaptation is proposed,which consists of an image translation module and two image denoising modules.Firstly,the image translation module is used to convert the synthetic infrared noise image to the real infrared noise image.Then,the converted image and the original image are used as the training data for the denoising module and the consistency loss function is used to make the two image denoising modules produce consistent results.Finally,the trained network framework is used for infrared image denoising task.The experiments show that the proposed algorithm has higher quantitative metrics and better visual quality on synthetic infrared noise datasets compared with BM3D,DnCNN and ADNet,and has the same excellent denoising effect on real infrared noise datasets.It is demonstrated that the algorithm has good generalization capabilities and is able to recover clear infrared images under real noise.