基于改进PCL模型的输电线路缺销螺栓弱监督检测方法

销子缺失是输电线路中常见的螺栓缺陷,及时检测出缺销螺栓对输电线路的安全运行至关重要。基于全监督检测模型的螺栓缺陷检测需要目标级标注,目标级标注会消耗大量的人力物力,为减少这种消耗,提出一种基于改进PCL(Proposal Cluster Learning)模型的输电线路缺销螺栓弱监督检测方法,仅利用图像级标注实现缺销螺栓检测。引入通道注意力机制,生成加权特征图,突出目标区域特征,有效地挖掘出螺栓的位置信息；采用加权交叉熵损失函数,控制正负样本对损失值的贡献,增大困难样本的损失比重,提高模型对螺栓目标的关注程度和识别能力；融合全监督的多任务学习思想,使模型能随着迭代次数的增加修正预先得到的边界框。实验结果表明,在测试集上,相比于基础模型,改进后的模型缺销螺栓的AP(Average Precision)值提升了25.6%,mAP(mean Average Precision)值提升了25.4%,最终验证了本文方法的鲁棒性。

Weakly Supervised Detection Method for Pin-missing Bolt of Transmission Line Based on Improved PCL Model

Pin missing is a common bolt defect in transmission line. It is important for the safe operation of transmission line to detect pin-missing bolt in time. The bolt defect detection based on fully supervised detection model needs object-level labeling, which will consume much human and material resources. In order to reduce this consumption, a weakly supervised detection method for pin-missing bolt of transmission line based on improved PCL(Proposal Cluster Learning) model is proposed, which only uses image-level labeling to detect pin-missing bolt. The channel attention mechanism was introduced to generate weighted feature map, highlight the features of target area, and effectively mine the bolt position information. The weighted cross entropy loss function was used to control the contribution of positive and negative samples to the loss value, increase the loss proportion of difficult samples, and improve the attention and recognition ability of the model to the bolt target. The fully supervised multi-task learning idea was integrated, so that the model could revise the pre-obtained bounding box with the increasement of iteration. The experimental results show that on the test set, compared with the basic model, the AP(Average Precision) value of pin-missing bolt of the improved model is increased by 25.6% and the mAP(mean Average Precision) value of is increased by 25.4% , which verifies the robustness of the proposed method in the paper.