Advances in Deep Learning-Based Ore Particle Size Detection: A Review of Methods, Challenges, and Trends

Ore particle size is a fundamental parameter in mineral processing, directly affecting grinding efficiency, equipment performance, and downstream product quality. Traditional manual and mechanical measurement methods are time-consuming, low in accuracy, and unsuitable for continuous monitoring. Recently, deep learning has emerged as a promising solution for automated ore size detection. This review systematically introduces deep learning methods for ore particle analysis, with a focus on two major paradigms: object detection (including anchor-based and anchor-free models) and image segmentation (including semantic segmentation, instance segmentation, and boundary regression). The performance of each method is compared across varying ore stacking scenarios, such as heavy occlusion, irregular particle shapes, and dusty environments, with an emphasis on their respective strengths and limitations. In addition, the review identifies major technical, equipment-related, and data-centric challenges that impede industrial deployment. These challenges include the development of robust algorithms, ensuring reliable real-time operation under adverse conditions, and securing high-quality annotated datasets. Recent advancements are examined, including weak supervision, few-shot learning, and multimodal fusion of RGB (Red, Green, Blue), depth, and infrared data. To enable intelligent and scalable ore particle size monitoring systems, future efforts should focus on building accurate, efficient, and generalizable models supported by self-supervised pretraining and sensor integration.