Edge-Aware Transformer for Adhesion Object Segmentation in XRT-Based Ore Presorting

Jie Huang, Gaochang Wu, Jingxin Zhang, Tianyou Chai

Abstract: X-ray transmission (XRT)-based ore presorting is a key technology for achieving efficient, sustainable, and energy-saving mineral processing. However, in practical industrial applications, densely overlapping ores frequently result in edge adhesion, which poses significant challenges for existing intelligent ore presorting systems and ultimately degrades the ore grade. To overcome this limitation, we propose in this work a general pipeline for adhesion object segmentation. Specifically, we design an adhesion edge-aware transformer, dubbed AEformer, which uses an edge enhancement module to sharpen object edges and an edge-aware attention module to capture adhesion regions, while preserving fine-grained details. Meanwhile, we introduce a unique edge focal loss function to prioritize adhesion regions during training, considerably improving segmentation accuracy. To address the scarcity and high cost of annotating adhesion ores in XRT images, we also propose a human-in-the-loop labeling strategy based on a foundation model to improve dataset quality and efficiency, and we apply this strategy to generate adhesion ore segmentation datasets in XRT images. Industrial validation in XRT-based intelligent ore presorting equipment confirms that our method remarkably improves ore presorting performance. Furthermore, extensive evaluation of two public datasets with similar challenges demonstrates the generalization, highlighting its potential for broader deployment in other intelligent mechatronics systems.

External IDs:doi:10.1109/tmech.2025.3604612