Bibliographic Details
| Title: |
Efficient Detection Algorithm for Off-Size Ore Based on Improved YOLOv12. |
| Authors: |
Liu, Yuanwen1 2209194374@qq.com, Li, Zhengpeng1 lkdlzp0901@163.com, Zhang, Xiaolin1 zhangxiaolin@ustl.edu.cn, Yang, Bin1 yangbin673039297@126.com, Hu, Jun1 320083700074@ustl.edu.cn |
| Source: |
Engineering Letters. Aug2026, Vol. 34 Issue 8, p3579-3590. 12p. |
| Subject Terms: |
*Feature extraction, *Ore-dressing, *Convolutional neural networks, *Object recognition (Computer vision), *Deep learning |
| Abstract: |
Efficient mineral beneficiation depends critically on the precise, rapid screening of off-size ore. However, harsh industrial environments—characterized by dust interference and volatile lighting—pose formidable challenges to conventional deep learning architectures. These aforementioned challenges have been extensively addressed in existing literature. To address these challenges, this study offers two primary contributions. First, we construct and release the first high-quality benchmark dataset for off-size ore detection, bridging the data gap in this industrial scenario. Second, we propose an efficient ore detection model based on an improved YOLOv12, named YOLO-OreDet. The proposed YOLO-OreDet architecture strategically integrates a Coordinate Attention (CoordAtt) mechanism within the neck network’s feature fusion layer. This integration significantly augments the model’s spatial awareness, facilitating the precise localization of key features amidst complex backgrounds. To improve the model's adaptability to the varied morphologies and textures of ore, standard convolutional layers were replaced by Omni-Dimensional Dynamic Convolutions (ODConv), significantly augmenting its feature extraction efficacy. Experiments on our proprietary dataset show that YOLO-OreDet achieves a mAP@0.5 of 99.5% and mAP@0.5:0.95 of 97.7%. These represent improvements of 0.1 and 2.3 percentage points, respectively, over the baseline YOLOv12 model. These results validate the effectiveness and superiority of our proposed strategies for industrial off-size ore detection. [ABSTRACT FROM AUTHOR] |
| Database: |
Academic Search Index |