摘要
现有煤矿胶带运输异物检测方法检测精度较低、检测速度较慢,YOLOv3算法有较快的检测速度和较高的检测精度,但其用于煤矿胶带运输异物检测时存在对小目标检测效果不佳、容易出现漏检和正负样本不均衡等情况。针对上述问题,设计了Fast_YOLOv3算法:通过改进先验框及边界框,以适应煤矿胶带运输小目标异物检测场景;增加反卷积网络,以提高小目标异物检测能力;引入Focal Loss改进损失函数中负样本置信度的交叉熵,解决正负样本数量不均衡问题,提高检测精度。设计了StiPic数据增强方法,对煤矿胶带运输图像进行预处理,以提高Fast_YOLOv3模型训练效率及对小目标异物的检测精度。实验及现场测试结果表明,Fast_YOLOv3算法对于胶带运输异物的平均检测精度达90.12%,平均检测时间为35 ms,对小目标异物的检出率达93.50%,满足胶带运输现场对异物检测精度和实时性的要求。
The existing foreign object detection methods in coal mine belt transportation are low in detection precision and slow in detection speed,and YOLOv3 algorithm has faster detection speed and higher detection precision.However,when it is used in foreign object detection in coal mine belt transportation,there are problems such as poor detection effect on small targets,easy to appear missing detection and imbalance of positive and negative samples.In order to solve the above problems,Fast_YOLOv3 algorithm is designed.By improving the priori box and bounding box,the algorithm is adapted to the detection scenario of small target foreign object in coal mine belt transportation.By adding the deconvolution network,the algorithm is able to improve the detection capability of small target foreign object.By introducing the Focal Loss to improve the cross entropy of the negative sample confidence in the loss function,the algorithm is able to solve the problem of imbalance in the number of positive and negative samples so as to improve the detection precision.The StiPic data enhancement method is designed to preprocess the coal belt transportation image to improve the training efficiency of the Fast_YOLOv3 model and the detection precision of small target foreign objects.The experimental and field test results show that the Fast_YOLOv3 algorithm can detect foreign objects in the belt transportation with an average precision of 90.12%,an average detection time of 35 ms,and a detection rate of 93.50%for small target foreign objects,which meets the requirements of foreign objects detection precision and real-time detection in the belt transportation field.
作者
任国强
韩洪勇
李成江
尹燕芳
REN Guoqiang;HAN Hongyong;LI Chengjiang;YIN Yanfang(Department of Electrical Information, Shandong University of Science and Technology, Jinan 250031, China)
出处
《工矿自动化》
北大核心
2021年第12期128-133,共6页
Journal Of Mine Automation
基金
国家自然科学基金资助项目(61572300)
山东省高等学校科技计划资助项目(J18KA328)
山东省重点研发计划项目(2016GGX105013)
山东省安全生产科技项目(F2010-004)。
