Image processing, as an efficient and accurate technology, has been widely applied to characterize granular object morphology in many fields, such as construction engineering, material science, agriculture, etc. Traditional static image processing is not autonomous because it cannot automatically segment contacting particles. In contrast, the current deep-learning-based algorithms can achieve high degree of autonomy in instance segmentation given it is well trained. However, lack of training data is a common pain point as it requires extensive manual labour using the conventional labelling tools. In this study, using sand as an example, we proposed a mask labelling methodology that can establish a large and diverse training set without manual labelling. The trained Mask-RCNN demonstrates excellent performance on a densely packed particle image. Using the data labelling method proposed in this study and the deep-learning algorithms, fully automated image processing can be realized for granular materials without massive manual labelling workload.

图像处理作为一种高效、准确的技术，已广泛应用于建筑工程、材料科学、农业等许多领域的颗粒状物体形态表征。传统的静态图像处理不能自动分割接触的颗粒，因此不具有自主性。相比之下，当前基于深度学习的算法在训练有素的情况下可以在实例分割中实现高度自治。然而，缺乏训练数据是一个常见的痛点，因为它需要使用传统的标签工具进行大量的手工劳动。在这项研究中，以沙子为例，我们提出了一种掩码标记方法，该方法可以建立一个庞大而多样化的训练集，而无需手动标记。经过训练的 Mask-RCNN 在密集的粒子图像上表现出优异的性能。