An Improved Lightweight YOLOv5 Algorithm for Detecting Railway Catenary Hanging String

Abstract

Aiming at the problems of small target and low recognition accuracy of high-speed railway contact network hanging chord defects, this paper proposes a target detection algorithm for hanging chord defects based on YOLOv5. To enhance the original YOLOv5 algorithm, the MobielNetv3 module was used as the efficient and lightweight backbone feature extraction network. Depth-separable convolution was adopted instead of standard convolution, reducing the number of network parameters by $2\times 10 ^{6}$ and increasing detection speed by 23%. Introducing BiFPN feature pyramid structure with fusion of different feature layers in neck network improves detection accuracy by 0.4%. Adding CBAM attention mechanism at the prediction end improves the feature extraction ability of the model for small target images, which further improves the detection accuracy by 0.5%. The loss function CIoU was improved to Focal EIoU in order to solve the problems of unbalanced sample datasets and vanishing IoU gradients during the training process. The experimental results exhibit that the improved algorithm achieves an average accuracy of 98.5% on the dataset, a 39% enchancment in model detection speed and a 28% reduction in model parameters, verifying that the algorithm has the advantages of high recognition accuracy and fast detection speed. It can effectively solve the technical difficulties in the detection of defects in the existing contact network suspension chords, and provides a new way of thinking for intelligent railway inspection.

Keywords

• YOLOv5
• lightweight network
• image detection
• railway
• deep learning