【摘要】 对闭路电视数据进行城市排水管道缺陷识别处理，常需要人工或者基于深度学习的方法来完成。但人工识别工作量和耗时巨大，而传统的深度学习方法还存在复杂缺陷识别准确率低和需要大量标签等问题，无法满足排水管道异常的类别检测和精确位置确定的要求。为此，发展了一种交互式的基于深度学习的排水管道缺陷内窥检测智能识别方法，可以在样本量较少的情况下不断提高识别效果。该方法通过适当的人工干预，校正识别的结果，并将识别结果自动加入到样本库中，积累和丰富各种缺陷的样本；通过对不断更新的样本的持续学习，可以迅速提高缺陷的识别准确度。在武汉市排水管道视频数据的缺陷识别试验中，虽然各类缺陷样本总共只有1 627个，但总体缺陷识别准确率可以达到68%,验证了该方法的有效性，展示了该方法有良好的应用前景。更多还原

【Abstract】 Defect identification in urban drainage pipelines based on closed-circuit television(CCTV) data often requires laborious manual efforts or deep learning-based methods. However, manual identification is characterized by an enormous workload and is time-consuming, while traditional deep learning approaches suffer from complex defect recognition with low accuracy and the need for a significant amount of labeled data, which are inadequate for anomaly classification and precise defect localization in drainage pipelines. To address these challenges, this paper proposes an interactive deep learning-based method for intelligent identification of defects in drainage pipelines, enabling continuous improvement of recognition performance even with limited training samples. The approach leverages appropriate human interventions to correct identification results and automatically incorporates them into the sample repository, accumulating and enriching a diverse set of defect samples. Through continuous learning from updated samples, the accuracy of defect identification can be rapidly enhanced. The effectiveness of the proposed method is validated through an experiment with only 1 627 defect samples from video data of the drainage pipeline from Wuhan, China, and its identification accuracy of overall defects could reach 68 %, demonstrating promising application prospects.更多还原