Dynamic cracking along an interface of brittle material is a fundamental knowledge of rock faulting but remain largely enigmatic. Laboratory investigation of the extremely fast process requires the use of advanced high-speed photography, which limits wide participation in this research. This study applies image data augmentation and Convolutional Neural Network (CNN) to assist high-speed photography for detection of crack tip location, extending the capability of high-speed photography with a low frame rate to achieve a high prediction accuracy. The prediction accuracy generally increases with a smaller kernel size to produce more image slices for model training, but the kernel size exists a lower limit for reasonable prediction, which is 0.8 × 30 mm in our study. The selection of kernel size is particularly important for a less advanced camera with a frame rate lower than 70,000 frame-per-second. Additionally, the digital image correlation technique can assist the identification of crack tip location on low-resolution images to provide image slices with precise information of crack tip for model training. Robust CNN models can be built with multiple static and dynamic loadings, different image resolutions, and various heterogeneous rock materials and used to construct a graphic reference for researchers to evaluate the best performance of high-speed photography and machine learning-assisted performance improvement.

中文翻译：

---

在数据增强和机器学习的辅助下，使用高速摄影智能检测沿脆性材料界面的动态裂纹


沿脆性材料界面的动态开裂是岩石断层的基础知识，但在很大程度上仍然是个谜。对极快过程的实验室研究需要使用先进的高速摄影，这限制了这项研究的广泛参与。本研究应用图像数据增强和卷积神经网络（CNN）辅助高速摄影检测裂纹尖端位置，扩展了低帧率高速摄影的能力以实现高预测精度。预测精度通常随着内核尺寸的减小而增加，以产生更多的图像切片用于模型训练，但内核尺寸存在合理预测的下限，在我们的研究中为 0.8 × 30 mm。对于帧速率低于 70,000 帧/秒的不太先进的相机来说，内核大小的选择尤其重要。此外，数字图像相关技术可以辅助低分辨率图像上裂纹尖端位置的识别，为图像切片提供精确的裂纹尖端信息用于模型训练。可以使用多个静态和动态载荷、不同的图像分辨率和各种异质岩石材料来构建鲁棒的 CNN 模型，并用于构建图形参考，供研究人员评估高速摄影和机器学习辅助性能改进的最佳性能。