A Comparative Detection and Multitask Learning Framework for Multimodal Defect Data in Drainage Pipelines

Continuous and accurate structural health monitoring of drainage pipelines is of great significance for ensuring the safe operation and scientific maintenance of drainage systems. Optical defect data and sonar defect data are generated during drainage pipeline inspection. To address the challenge of automatically selecting pairable data samples from multimodal defect data and efficiently integrating them for training to achieve multitask defect detection, this paper proposes an effective data management strategy. Firstly, a single‐task detection framework is introduced, capable of achieving higher detection accuracy on both optical and sonar datasets, which is then used to develop a multimodal defect detection system for pairing defect data. Furthermore, a multitask detection framework, CerberusDet‐U, is proposed, enabling multitask joint detection without altering the structure of individual sonar datasets. Experimental results show that the proposed single‐task detection framework improves F1, mAP 0.5 , and mAP 0.5–0.95 on the sonar dataset by 2.3%, 1.7%, and 1.3%, respectively, compared to the baseline and by 2.7%, 3.6%, and 2.1% on the optical dataset. The proposed multitask detection framework achieves improvements of 2.8%, 3.6%, and 1.7% on the sonar_A dataset and 2.1%, 3.9%, and 0.9% on the sonar_B dataset, respectively, over the baseline. These optimization algorithms and engineering applications not only enhance the accuracy and efficiency of multimodal defect detection but also provide reliable data support and technical assurance for continuous structural health monitoring of drainage pipelines.