CBAM-Enhanced YOLOv5 for Automated Detection of Urban Underground Drainage Pipe Defects

Abstract

The timely detection and repair of defects and damages in underground drainage pipes are crucial for the normal operation of cities. Focusing on the defect detection and damage localization of urban underground drainage pipes, this paper introduced the Convolutional Block Attention Module (CBAM) to the You Only Look Once version 5 (YOLOv5) algorithm to enhance its ability of feature extraction. Then, several different loss functions were compared. Experimental analyses were carried out using the sewer-ML dataset. The results showed that among different versions of the model, the YOLOv5l model had better overall performance. Compared with the Squeeze-and-Excitation and coordinate attention modules, the CBAM had a better optimization effect on the YOLOv5 algorithm, bringing a 5.7% mean average precision improvement. The detection effect obtained when Softmax Intersection over Union (SIoU) was used as the loss function was better than efficient Intersection over Union (EIoU) and Focal EIoU. When CBAM and SIoU were used for optimization together, the improved YOLOv5 algorithm achieved a mean average precision of 93.37% and a frame rate of 85 frames per second, which had an advantage over the other algorithms. The method can be used in practice.