Optimizing Damage Detection in Pipelines with Drone-Based Deep Learning Models

Abstract

This study investigates the application of deep learning algorithms, specifically SSD and YOLOv7, for pipeline damage detection using remote sensing technology. The research evaluates the performance of both algorithms in terms of precision, recall, mean average precision (mAP), and F1 score, using a dataset collected through UAV-based imagery. SSD demonstrated superior precision but slightly lower recall, while YOLOv7 excelled in recall and overall detection ability, making it more suitable for comprehensive inspections where missing defects is critical. The findings emphasize the importance of context-specific algorithm selection, with SSD being ideal for real-time monitoring systems and YOLOv7 for applications requiring high recall. Furthermore, the study explores potential improvements to SSD, including transfer learning, data augmentation, and advanced feature extraction techniques, to enhance recall and overall performance. The results offer valuable insights for optimizing pipeline damage detection in varying operational environments.