A transfer learning-based YOLO network for sewer defect detection in comparison to classic object detection methods

Abstract

Deep learning has shown promising performance in automated sewer defect detection, however, is generally data-driven and computationally intensive. Transfer learning (TL) solves the problem of data limitations and avoids the need to build models from scratch. This study compared the performance of a TL-based YOLO network (with 11 pretrained backbone CNNs) with four mainstream object detection methods (ODMs) for detecting five types of sewer defects. Results showed that the transferred YOLO methods generally outperformed the other ODMs, with improved detection precision, computation speed and intersection over union (IoU). Among the CNNs, Resnet18 achieved the best performance, while Inceptionresnetv2 was the least effective. The ODMs worked best in detecting disjoint, whereas tree root and crack were most challenging to predict. The work not only illustrated the benefits of TL, but also provided technical guidance to practitioners who lack expertise in ODMs and rely on TL for better sewer defect detection.

Keywords

• Object detection
• Transfer learning
• Pretrained CNNs
• YOLO network
• Sewer defects