A Novel Method for Particle Instance Segmentation and Size Measurement
Fei Li, Xiaoyan Liu, Yufeng Yin, Zongping Li
Abstract
Deep learning has demonstrated significant potential in particle size measurement across various fields, including mining, agriculture, and cell processing. However, its performance heavily relies on large datasets with accurate particle annotations, which are costly and time consuming to acquire. Additionally, deep learning models suffer from performance degradation when confronted with particle appearance variations, consequently affecting particle segmentation and size measurement. In this article, we propose a novel method to generate a synthetic dataset of particles and automatically annotate them using a “Copy-and-Paste” technique. This method allows us to create a large and diverse particle dataset without manual annotation. Concurrently, we adapt and improve the instance segmentation model, Mask R-CNN, to develop a model specifically designed for particle instance segmentation. We train the improved Mask R-CNN with the synthetic dataset and test it on synthetic and real datasets. The experimental results show that our method can segment different types of real particles, such as iron ore, pebbles, and aggregates, with high accuracy. We also measure and analyze the size distribution of the segmented real particles, which agree well with manual sieving results, with errors within 7% for different types of real particles. Our dataset and code are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/lifeiwen/</uri> Auto-Label-For-Ore-Instance-Segmentation.