基于双分支串行混合注意力的输电线路缺陷检测深度神经网络模型

Deep Neural Network Model for Transmission Line Defect Detection Based on Dual-branch Sequential Mixed Attention

检测输电线路缺陷并及时维修可以确保电网的安全稳定,具有重大的实际意义。但输电线路图像背景复杂、元件尺寸小,导致现有的目标检测模型不能取得很好的效果,因此文中提出了基于双分支串行混合注意力的输电线路缺陷检测深度神经网络模型。该模型设计了DBSA(Dual-branch Serial Attention)双分支串行混合注意力,从而将更多的权重放在缺陷上,并提出了WCFPN(Well-connected Feature Pyramid Network)特征金字塔,让经DBSA提取的特征充分融合,从而增强模型检测小目标的能力。DBSA将特征图沿高度和宽度两个分支压缩并用一维卷积提取注意力,WCFPN设计了一种包含跨尺度融合和跳层连接的新型融合路径,让经DBSA提取的高层语义信息和低层空间信息进行更充分的交互。最后在绝缘子自爆、防振锤损坏、鸟巢异物、水泥杆破损和输电线路缺陷5个数据集上进行实验,结果显示所提模型取得了最佳的检测效果,在5个数据集上的平均AP50和AP分别为84.3%和46.1%,相比目前最先进的模型YOLOv7分别提升了3.7%和3%。

Detecting defects in transmission lines and repairing them in a timely manner is of great practical significance for ensuring the safety and stability of the power grid.However,due to the complex background and small component size of transmission line images,existing object detection models cannot achieve satisfactory results.Therefore,this paper proposes a deep neural network model for detecting defects in transmission lines based on dual-branch serial attention.The model designs dual-branch serial attention(DBSA)to allow the model to focus more weight on the defects,and proposes well-connected feature pyramid network(WCFPN)to enable full fusion of the features extracted by DBSA,thereby improving the model's ability to detect small targets.DBSA compresses the feature map along the height and width branches and extracts attention using one-dimensional convolution to achieve fine-grained control over the features.WCFPN designs a new fusion path that includes cross-scale fusion and skip-layer connections,allowing high-level semantic information and low-level spatial information extracted by DBSA to interact more fully.Finally,experiments are conducted on five transmission line datasets,including insulator explosion,damaged anti-vibration hammer,bird's nest debris,broken cementpole and transmission line defect,and the proposed model achieves the best detection performance.The average AP 50 and AP of the five datasets is 84.3%and 46.1%,respectively,which is 3.7%and 3%higher than that of the state-of-the-art model YOLOv7.