.png){kind=icon}
.png){kind=icon}
Return
Detection and Effectiveness of Improved Drainage Pipe Defects-Based Semantic Segmentation Labeling Methods
.png){kind=icon}
PRE
AI
DOI：10.1061/JPSEA2.PSENG-1898.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This study investigates how to optimize and improve defect sample labeling techniques using 15 different types of drainage pipe flaws. The link between the overall labeling approach and model performance is not fully taken into account by traditional sample labeling methods. To achieve this, this study suggests a thorough improvement plan that addresses three areas: labeling method optimization, identifying the ideal range of labeling accuracy, and developing labeling principles. First, the study refines the defect labeling technique, examines and contrasts the labeling techniques of various defect categories, identifies the best labeling technique for each defect category, and combines them to create the labeling optimization strategy. Second, this paper identifies the ideal range of labeling accuracy by analyzing the effects of varying labeling accuracy on model performance. To increase the labeling's accuracy even more, the labeling guidelines that apply to drainage pipe samples that are defective are finally requested. According to the experimental results, the sample set built using this optimization strategy is used for model judgment with the mAP value of 72.8% and precision and recall rates of 84.0% and 63.7%, respectively, which makes the semantic segmentation model judgment with good results without enhancing the model. The study demonstrates that the deep learning model's judgment performance can be improved by optimizing the sample labeling strategy, which has both theoretical and practical implications.
Keywords:
Drain defect detection
Sample library
Sample labeling
YOLO v8
Ablation experiment
Journal
J
IF:
1.7
Papers:
69
Citations:
1.2K