为提高排水管道缺陷的AI检测模型效果,研究一种基于小样本Faster R-CNN算法的半监督学习方法.该方法利用无缺陷正常管道图像,随机为其生成伪标签,并将其混入具有缺陷标签的图像集中,实现半监督式训练.对管道CCTV视频进行了检测实验,对比只采用有缺陷图像集训练,混合图像集训练的 AI 模型对各种管道缺陷的检出率都有较高水准,且缺陷误检率从34.85%下降至6.12%.研究结果表明,采用混合图像集能够影响AI检测模型的缺陷识别效果;在小样本管道缺陷图像训练集和验证集中额外混入十分之一的伪标签图像,训练出的检测模型具有最佳的平均精度均值.
Study on the Effects of Mixed Image Dataset on AI Detection of Drainage Pipeline Defects
In order to enhance the effectiveness of the AI detection model for identifying defects in drainage pipelines,a semi-supervised learning approach based on Faster R-CNN algorithm has been proposed for the small-sample image dataset.In this approach,the defect-free pipeline images are randomly genera-ted with pseudo-labels,and mixed into the image dataset with defect annotations to facilitate semi-su-pervised training.Detection experiments have been conducted on actual drainage pipeline CCTV videos,and compared with models trained on image datasets without pseudo-labels.The results demonstrate that the AI model trained on mixed image dataset achieves high detection rates for various pipeline defects,and the false detection rate is reduced from 34.85%to 6.12%.It is also shown that learning from different mixed image datasets significantly influences the defect recognition capability of AI detection models.For the small-sample image dataset,augmenting one-tenth more pseudo-labeled images into the training and validation sets results in the detection model attaining the best mean average precision.
NETL
NSTL
万方数据
