| 摘 要: 为应对复杂环境下城市排水管道缺陷检测中普遍存在的多尺度缺陷漏检、误检以及特征提取困难等挑战,提出一种基于YOLOv8n改进的检测算法。首先,引入SPPFCSPC模块重构多尺度特征提取,减少计算冗余；其次,在主干末端集成SKA注意力机制以融合多尺度特征并聚焦关键区域；最后,在检测头前加入SWS空间注意力模块,增强小目标特征表达。实验结果显示,模型保持原有检测速率的同时检测准确率达到83%,较基准模型提升2.8%,漏检率与误检率分别降低1%和2%。在公开数据集上的泛化实验进一步证实了该算法的有效性。 |
| 关键词: 地下排水管道 缺陷检测 深度学习 多尺度缺陷 注意力机制 |
|
中图分类号:
文献标识码:
|
|
| Improved Underground Drainage Pipeline Defect Detection Algorithm Based on YOLOv8 |
|
TangMingping, DongQin, Wangwei
|
Yancheng Institute of Technology
|
| Abstract: To tackle the widespread challenges of multi-scale defect omission, false detection, and difficult feature extraction in urban drainage-pipeline inspection under complex environments, we propose an improved detector built upon YOLOv8n. First, an SPPFCSPC module is introduced to restructure multi-scale feature extraction and cut computational redundancy. Second, an SKA attention mechanism is embedded at the backbone’s terminus to fuse multi-scale features and spotlight critical regions. Finally, an SWS spatial-attention block is inserted ahead of the detection head to strengthen small-target representation. Experiments show the model achieves an 83 % mean average precision, 2.8 % higher than the baseline, while lowering the miss-detection and false-detection rates by 1 % and 2 %, respectively, without sacrificing the original inference speed. Generalization tests on a public dataset further confirm the algorithm’s effectiveness. |
| Keywords: underground drainage pipeline defect detection deep learning multi-scale defects attention mechanism |
