基于自监督学习方法SwAV实现煤矿场景目标检测

Realization of Coal Mine SceneTarget Detection based on Self-Supervised Learning Method SwAV

随着智能矿山建设的推进,基于目标检测方法的智能视频监控系统在矿山生产采掘运和生产环节得到大量应用。其中目标检测方法常采用监督学习方法,其存在标注数据成本大等问题。利用无监督学习方法SwAV在包括调度室内的各个场景视频图片进行预训练,并迁移到检测方法上进行目标检测,实现煤矿调度室空岗、停产煤矿在生产等监测功能。实验证实SwAV在上游预训练和下游目标检测上采用同一分布,能显著提升对下游目标检测任务性能。

With the advance of intelligent mine construction,intelligent video monitoring system based on target detection method has been widely used in mining,transportation and production.Supervised learning method is often used in target detection,which has some limitations,such as high labeling data cost.The unsupervised learning method SwAV is used to pretrain the video images of various scenes including the dispatching room,and migrate to the target detection method,so as to realize the monitoring functions of empty post in the coal mine dispatching room and production of discontinued coal mines.Experiments show that SwAV can significantly improve the performance of downstream target detection tasks by using the same distribution in upstream pretraining and downstream target detection.