Deep neural networks and deep learning have achieved great success in many signal and image processing applications, especially those with large-scale annotated training data for supervised learning. Although in principle deep-learning methods can be applied to boost the performance of processing materials-science images, i.e., microscopic images that capture important microstructures of various material samples, many priors and requirements in materials science must be considered to maximize performance gains. In this article, we focus on the important problem of detecting objects of interest from microscopic materials-science images and introduce different approaches to incorporate several such priors, including object shape, symmetry, and 3D consistency, into deep learning to tackle this problem. In particular, we explore the use of these three priors to enable network training with fewer data annotations, which is highly desired in materials science. This tutorial-style article will summarize contributions in the literature as well as our current research achievements, and we hope it can provide an initial insight to new researchers who are interested in using deep learning for materials-science image processing.

中文翻译：

---

材料科学图像中物体检测的深度学习：教程


深度神经网络和深度学习在许多信号和图像处理应用中取得了巨大成功，特别是那些用于监督学习的大规模带注释训练数据的应用。虽然原则上深度学习方法可用于提高处理材料科学图像（即捕获各种材料样本的重要微观结构的显微图像）的性能，但必须考虑材料科学中的许多先验和要求，以最大限度地提高性能增益。在本文中，我们重点关注从微观材料科学图像中检测感兴趣对象的重要问题，并介绍了将多个此类先验（包括对象形状、对称性和 3D 一致性）融入深度学习中以解决此问题的不同方法。特别是，我们探索使用这三个先验来实现用更少的数据注释进行网络训练，这在材料科学中是非常需要的。这篇教程式的文章将总结文献中的贡献以及我们当前的研究成果，我们希望它能为有兴趣使用深度学习进行材料科学图像处理的新研究人员提供初步见解。