[1]关子皓,张智豪,何淑雅,等.基于YOLO和特征引导SAM的管道缺陷检测方法[J].中国给水排水,2026,42(11):122-127.
GUANZihao,ZHANGZhihao,HEShuya,et al.Drainage Pipeline Defect Detection via YOLO and Feature Extraction Guided SAM Segmentation[J].China Water & Wastewater,2026,42(11):122-127.
点击复制
GUANZihao,ZHANGZhihao,HEShuya,et al.Drainage Pipeline Defect Detection via YOLO and Feature Extraction Guided SAM Segmentation[J].China Water & Wastewater,2026,42(11):122-127.
基于YOLO和特征引导SAM的管道缺陷检测方法
中国给水排水[ISSN:1000-4062/CN:12-1073/TU]
卷:
第42卷
期数:
2026年第11期
页码:
122-127
栏目:
出版日期:
2026-06-01
- Title:
- Drainage Pipeline Defect Detection via YOLO and Feature Extraction Guided SAM Segmentation
- Keywords:
- pipeline defect detection; SAM model; phase consistency; semantic segmentation; defect grading
- 摘要:
- 在城市排水系统日益复杂的背景下，传统人工巡检和基于单一模型的缺陷检测方法面临着识别效率低、分割精度不足以及依赖大量标注样本等局限性。为应对这些技术挑战，提出了一种融合YOLOv4目标检测算法与多特征提取引导SAM（segment anything model）的缺陷检测创新方法。实验表明，虽然YOLO + SAM融合方法在常规场景下能实现有效分割，但在处理复杂缺陷形态（如裂缝、树根侵入和错口等）时，其分割性能仍存在不足。为此，进一步引入相位一致性分析、分形维数矩阵计算和边缘检测优化三种多模态特征提取方法，以提升模型对异质缺陷的特征表达能力。结果显示，YOLO +特征提取+SAM的组合方案在裂缝、树根侵入和错口等典型缺陷检测中，相较于YOLO+SAM方案取得了显著提升：召回率提高了17.1%～24.3%，精确率提升了24.9%～44.4%，F1分数增加了23.1%～32.4%。该方法不仅实现了更清晰的缺陷分割边界和更高的识别精度，还大幅降低了模型训练成本和标注需求，具有重要的工程应用价值。
- Abstract:
- With the increasing complexity of urban drainage systems, traditional manual inspection and single model-based defect detection methods suffer from low recognition efficiency, insufficient segmentation accuracy, and heavy reliance on large annotated datasets. To address these technical challenges, this study proposes an innovative defect detection method that integrates the YOLOv4 object detection algorithm with multi-feature extraction-guided SAM (segment anything model). The experiments demonstrated that, although the YOLO+SAM integrated approach achieved effective segmentation in conventional scenarios, its performance in handling complex defect shapes (such as cracks, root invasions, and disjoint) remained inadequate. To this end, three multimodal feature extraction methods—phase consistency analysis, fractal dimension matrix computation, and edge detection optimization—were introduced to enhance the model’s ability to represent heterogeneous defects. Experimental results showed that the combination of YOLO+feature extraction+SAM achieved significant improvements in detecting typical defects like cracks, root invasions, and disjoint compared to the YOLO+SAM approach: the recall rate increased by 17.1% to 24.3%, precision improved by 24.9% to 44.4%, and the F1 score increased by 23.1% to 32.4%. This method not only provides clearer defect segmentation boundaries and higher recognition accuracy but also significantly reduces model training costs and labeling requirements, demonstrating substantial engineering application value.
更新日期/Last Update:
2026-06-01