【摘要】 防震锤和线夹是输电线路自动化设备巡检过程中的重要巡检对象。针对两类目标检测时的多角度需求与样本量较少的问题,提出基于迁移学习的改进训练方法训练YOLOv3模型,降低训练时模型对防震锤与线夹样本的需求,并提高模型最终的准确性与泛化能力。首先,从不同时段、不同角度、不同季节、不同背景对防震锤和线夹线上目标开展了数据采集工作。其次,通过分析YOLOv3网络的卷积层结构,构建了多组迁移层数不同的训练模型,并在自主采集的防震锤、线夹小样本库上进行训练,之后通过比较分析,得到了最适合该小样本库的迁移学习模型。最后基于对模型实际检测图像的比较与分析,评估了通过迁移学习方法降低模型在防震锤、线夹小样本库上的训练成本的可行性。实验结果表明,通过迁移学习方法在该小样本库上训练YOLOv3网络,并在特征图等效输入层为31层时,模型的性能最好,此时模型的收敛速度比无迁移学习时提高了一倍,模型的平均精度均值(mAP)值提高了6.58%。其中防震锤单项AP值最高达到92.22%,比同类机器学习算法提高了近15%。更多还原

【Abstract】 Damper and clamp are important inspection objects in the transmission line automation inspection. In order to solve the problem of multi-angle requirements and small sample size when detecting two types of targets,an improved training method based on transfer learning was used to train the YOLOv3(You Only Look Once version 3)model. First of all,data collection work was carried out from different periods,different angles,different seasons,and different backgrounds. Secondly,by analyzing the structure of the YOLOv3,four training models with different transfer layers were constructed,and the training was performed. Finally,based on the comparison and analysis of the ture detection images,the feasibility of using the transfer learning method to reduce the training cost on damper and clamp dataset was evaluated. The experimental results show that the model performance is the best with 31 equivalent input layers of the feature map,while the convergence rate of the model is doubled that without transfer learning,the mean Average Precision(mAP)value of the model is increased by 6. 58%;the maximum value of the damper AP(Average Precision)is 92. 22%,and nearly 15%higher than that of the same machine learning algorithm.更多还原