Ede Jeffrey M, Beanland Richard
Department of Physics, University of Warwick, Coventry, England CV4 7AL, United Kingdom.
Ultramicroscopy. 2019 Jul;202:18-25. doi: 10.1016/j.ultramic.2019.03.017. Epub 2019 Mar 26.
We present an atrous convolutional encoder-decoder trained to denoise electron micrographs. It consists of a modified Xception backbone, atrous convoltional spatial pyramid pooling module and a multi-stage decoder. Our neural network was trained end-to-end using 512 × 512 micrographs created from a large dataset of high-dose ( > 2500 counts per pixel) micrographs with added Poisson noise to emulate low-dose ( ≪ 300 counts per pixel) data. It was then fine-tuned for high dose data (200-2500 counts per pixel). Its performance is benchmarked against bilateral, Gaussian, median, total variation, wavelet, and Wiener restoration methods with their default parameters. Our network outperforms their best mean squared error and structural similarity index performances by 24.6% and 9.6% for low doses and by 43.7% and 5.5% for high doses. In both cases, our network's mean squared error has the lowest variance. Source code and links to our high-quality dataset and pre-trained models are available at https://github.com/Jeffrey-Ede/Electron-Micrograph-Denoiser.
我们展示了一种经过训练用于对电子显微镜图像进行去噪的空洞卷积编码器-解码器。它由一个经过修改的Xception主干、空洞卷积空间金字塔池化模块和一个多级解码器组成。我们的神经网络使用从一个包含高剂量（每像素>2500计数）显微镜图像的大型数据集中创建的512×512显微镜图像进行端到端训练，并添加泊松噪声以模拟低剂量（≪300计数每像素）数据。然后对其进行高剂量数据（200 - 2500计数每像素）的微调。其性能以具有默认参数的双边、高斯、中值、总变差、小波和维纳恢复方法为基准进行评估。对于低剂量情况，我们的网络在最佳均方误差和结构相似性指数性能方面分别比它们高出24.6%和9.6%；对于高剂量情况，分别高出43.7%和5.5%。在这两种情况下，我们网络的均方误差具有最低的方差。源代码以及我们高质量数据集和预训练模型的链接可在https://github.com/Jeffrey-Ede/Electron-Micrograph-Denoiser获取。
