摘要
按照实际工况在实验室搭建煤矸分选平台,采用深层目标检测网络对煤矸石进行在线识别,根据分选时煤矸石的形状和大小,将目标检测网络中的特征金字塔设定为3个尺度,并确定锚(参考边界框)的形状和大小;比较10个IOU(Intersection-over-union)阈值下验证集的平均精度(AP),并在煤矸石分选平台对目标检测网络进行动态测试。结果表明:IOU为0.8时,目标检测网络的分类和定位效果最佳,动态识别的精确度和召回率均达到95%以上。
A platform of coal-gangue separation was set up in the laboratory according to the actual working conditions.Adopting object detection network to identify coal and gangue online and according to the shape and size of coal gangue during sorting,the feature pyramid in the object detection network was set as three scales,and the shape and size of the anchor were determined.Comparing the AP(average precision)of the validation set under 10 IOU(intersection over union)thresholds,and a dynamic test was carried out on the separation platform built.The results show that the classification and positioning effect of the object detection network was the best when the IOUwas 0.8,and the precision and recall of dynamic identification can reach more than 95%.
作者
高新宇
李博
王璐瑶
李廉洁
王学文
GAO Xinyu;LI Bo;WANG Luyao;LI Lianjie;WANG Xuewen(College of Mechanical and Vehicle Engineering,Shanxi Provincial Key Laboratory of Coal Mining Equipment,Taiyuan University of Technology,Taiyuan 030024,China)
出处
《中国粉体技术》
CAS
CSCD
2021年第4期77-83,共7页
China Powder Science and Technology
基金
山西省重点研发计划项目,编号:201903D121074。
关键词
煤矸分选
图像处理
深度学习
目标检测
coal-gangue separation
image processing
deep learning
object detection
