基于卷积块注意力模块和双向特征金字塔网络的接触网支持装置检测方法研究

Study on Detection Method of Catenary Support Device Based on Convolutional Block Attention Module and Bidirectional Feature Pyramid Network

接触网支持装置是接触网悬挂状态检测监测图像分析的关键对象,对支持装置零部件的检测定位是实现缺陷自动分析的基础。针对接触网支持装置零部件种类多、尺寸差异大、存在遮挡、部分零部件相似度高等问题,提出一种融合卷积块注意力模块(convolutional block attention module,CBAM)和双向特征金字塔网络(bidirectional feature pyramid network,BiFPN)的接触网支持装置检测方法。在YOLO v5s网络模型基础上,该方法通过CBAM增强接触网支持装置的特征提取,结合BiFPN,实现不同零部件分辨率特征图的融合。利用4C装置获得的图像数据集,开展验证试验。试验结果表明,相对YOLO v5s网络模型,融合CBAM和BiFPN的接触网支持装置检测方法,网络平均精度mAP@0.5提高2.12%;能显著提升小目标检测效果,提高定位的准确性和稳定性,对接触网状态的智能分析有重要意义。

The catenary support device is the key object in the analysis of suspension state of the catenary detection and monitoring images.Detection and positioning of the support device partps is the basis of automatic defect analysis.Aiming at the problems of many types of the catenary support device parts,large size difference,occlusion and high sim⁃ilarity of some parts,a detection method for the catenary support device combining convolutional block attention module(CBAM)and bidirectional feature pyramid network(BiFPN)was proposed.On the basis of the YOLO v5s net⁃work model,the method enhance the feature extraction of the catenary support device by CBAM,and BiFPN was com⁃bined to realize the fusion of feature maps of different parts.Validation tests were carried out using the image data set obtained by the 4C device.The test results show that,compared with YOLO v5s network model,the average accuracy of the network mAP@0.5 is improved by 2.12%by detection method combining the CBAM and BiFPN for the catena⁃ry support device.The detection method combining CBAM and BiFPN for the catenary support device can significant⁃ly improve the effect of small target detection,improve the accuracy and stability of positioning,which is of great sig⁃nificance for the intelligent analysis of catenary status.