Intelligent Detection of Drainage Pipeline Defects Based on Cross-Frame Annotation and Recall Optimization
Drainage pipeline defect detection predominantly relies on closed-circuit television (CCTV) inspection, which is labor-intensive, inefficient, and prone to missed detections. Although deep learning-based object detection has been applied, it suffers from low precision, recall, and speed in practical scenarios. This study proposes an engineering-oriented detection scheme achieving high recall and low miss rates. The annotation phase employs a cross-frame strategy combining manual labeling of first and last frames with interpolation and tracking-based refinement. Data preprocessing introduces perceptual hashing to identify similar images, enhancing training efficiency. For detection, a Faster R-CNN model is enhanced with Focal Loss to focus on hard examples, defect classification and grading, and a dynamic threshold strategy to improve recall. Validated on 5,068.72 m of real pipeline data, the method achieves a recall rate exceeding 98% across 16 defect categories, a miss rate of only 2% for grade 4 defects, and a 425% improvement in per-segment detection efficiency compared to manual screening. These results demonstrate the method's effectiveness in balancing recall, miss rate, and speed for engineering deployment.