基于YOLOv5n_Faster的煤矿变电所开关柜目标检测

Object Detection of Coal Mine Substation Switch Cabinets Based on YOLOv5n_Faster

为了满足煤矿变电所实时巡检的需求,需要使用速度快且精度高的目标检测算法。为此本文提出一种改进的YOLOv5n模型,使用PConv代替普通卷积,减少特征图冗余,提高计算速度。使用α-CIoU替换了YOLOv5中原本的CIoU损失函数,通过调整α参数的值来提高模型的自适应性能。为了解决自制数据集数据量不够的问题,使用图像增强对有限的数据集进行扩充,对开关等小目标进行额外图像增强。在对煤矿变电所开关柜数据集中的实测结果表明,使用改进后的YOLOv5n模型进行目标检测,该模型的平均准确度为98.8%,比原始的YOLOv5n模型提高了0.3%的精度,平均时间缩短了55.621 ms,可以同时满足目标检测的精度和速度需求,可应用于实际的开关柜检测场景。

To meet the real-time inspection needs of coal mine substations, a fast and accurate target detection algorithm is required. This paper proposes an improved YOLOv5n model that uses PConv instead of ordinary convolution to reduce feature map redundancy and improve computational speed. The α-CIoU loss function is used to replace the original CIoU loss function in YOLOv5, and the adaptive performance of the model is improved by adjusting the value of the α parameter. To address the is-sue of insufficient data volume in the self-made dataset, image augmentation is used to expand the limited dataset, with additional image augmentation applied to small targets such as switches. The experimental results on the switch cabinet dataset of a coal mine substation show that the im-proved YOLOv5n model achieves an average accuracy of 98.8% in target detection, which is 0.3% higher than the original YOLOv5n model, and reduces the average time by 55.621 ms. This model can simultaneously meet the accuracy and speed requirements of target detection and can be ap-plied in actual switch cabinet detection scenarios.