Aiming at the problems of complex application environment of overhead contact system,small size of steady ears,special installation direction,difficult identification,and poor performance of traditional target detection algo-rithms,a new method for detecting defects in steady ears of overhead contact system was proposed based on multi-scale feature fusion dense connectivity network(SSD)model.Firstly,the feature extraction network of SSD model was im-proved by combining DenseNet and Inception modules to share the context information among specific feature layers to enhance the feature extraction capability.Then the multi-scale feature maps for SSD detection were fused using FPN step by step from deep to shallow layers of the network to design a dense connected network model for multi-feature fusion.Fi-nally,fGIoU was used as the edge loss function to optimize the overlap between the real and predicted frames during train-ing.The collected image dataset of steady ears of a section of overhead contact system was detected and recognized.The results show that the steady ear defect detection method designed in this paper can detect the steady ear detachment and loosening under the complex backgrounds of the overhead contact system with strong robustness in images with different angles and brightness.
关键词
接触网/缺陷识别/定位线夹/特征金字塔/边界框损失函数
Key words
overhead contact system/defect identification/steady ear/feature pyramid/bounding box loss function
维普
万方数据