| description abstract | The water supply pipe system is an important component of the municipal pipe system. However, water supply pipes usually suffer from various defects, such as deposits and infiltrations, which severely affect their performance and result in millions of dollars being wasted on maintenance work. Therefore, timely and effective inspection of water supply pipes is very important. In recent years, automatic detection based on deep learning methods has had the advantages of high efficiency, low cost, and time saving, thus gradually replacing manual inspection for defects in the pipe system. To solve the problem of unclear image acquisition for water supply pipes, this paper proposes a novel automated detection method for water supply pipe defects, mainly involving the use of underwater image enhancement (UIE) algorithms to improve data set image quality, and an attention mechanism was utilized to improve the You Only Look Once X (YOLOX) model for defects detection. Experimental results demonstrate that the improved YOLOX model based on the data set enhanced by underwater image enhancement and attention mechanism achieved an average accuracy [mean average precision (mAP)] value of 92.4% and F1 score of 0.86, which are better than traditional models. Finally, an efficient and accurate automated detection procedure for water supply pipe defects was provided. The automatic detection method of water supply pipe defects constructed in this research has the following three significant practical advantages: (1) the UIE algorithm is applied to the water supply pipe image, which improves the quality and quantity of the data set; (2) the method combines the improved UIE data set and attention mechanism to promote the efficiency and accuracy of the object detection model; (3) the research provides a novel procedure for the automatic detection work of water supply pipe defects. New attempts have been made in three aspects—the use of underwater pipeline robots for detection, the establishment of a high-quality data set, and the training or prediction of the object detection model for water supply pipe—and good results have been achieved. For these reasons, this method can greatly reduce the workload of construction personnel, and effectively avoid the occurrence of detection error events caused by the misjudgment of technicians. | |
