YOLOv8-ORE: An Efficient Ore Segmentation Network based on Adaptive Feature Extraction and Attention-Enhanced Spatial Fusion

Yidan Long, Baoning Cai, Jianming Hu, Wei Hu, Wen Yang, Wenfeng Zhang, Qibing Qin

Abstract: Ore particle size is crucial for ore crushing quality and the mining industry. Improving ore particle segmentation accuracy is essential for reliable statistics, especially in complex scenarios. YOLOv8-ORE, a novel ore image segmentation model, addresses challenges such as diverse ores, dust, lighting conditions, and adhesion between particles. Firstly, to enhance small particle detection, Flexible C2f integrates Adaptive Kernel Convolution, effectively capturing multi-scale features and significantly improving the segmentation of tiny, irregular particles. Secondly, to improve edge detection in low-light conditions, Attention-Enhanced Spatial Pyramid Pooling, inspired by Efficient Layer Aggregation Networks’ gradient path design and partial self-attention, was designed to better distinguish particles from conveyor belt backgrounds in dim environments. Thirdly, to refine segmentation accuracy in blurry images, BiFPN was employed to enhance accuracy through an efficient weighted fusion mechanism. Moreover, the OreSegDataset(OSD) was constructed in this paper to support ore segmentation community development. OSD consists of 718 annotated ore images with 11,084 instances. Experimental results on OSD demonstrate that the proposed YOLOv8-ORE outperforms state-of-the-art methods, such as YOLOv8-seg. The OSD dataset is publicly available at: https://github.com/CQNU-ZhangLab/.

External IDs:dblp:journals/sivp/LongCHHYZQ25