A Dual Task Learning Framework for Multitype Bridge Condition Assessment Using UAV Imagery

Abstract Unmanned aerial vehicle (UAV) imagery combined with artificial intelligence offers unprecedented potential to improve bridge health monitoring, reducing reliance on labor-intensive and potentially hazardous manual inspections. However, the diversity of bridge types and damage categories poses significant challenges for existing convolutional neural network (CNN)-based methods, which often focus on single damage types or specific bridge structures, limiting their generalization and comprehensive applicability. To address these limitations, this study proposes a novel dual task deep learning framework, termed bridge damage detection network (BDDNet), for simultaneous detection and segmentation of multiple surface damages across various bridge types using UAV imagery. BDDNet integrates a global attention module (GAM) within its backbone to capture long-range dependencies and enhance multiscale feature representation. Additionally, a novel APC2f module based on pinwheel convolution is incorporated into the detection branch, enlarging the receptive field and refining feature learning for small and complex damages. Extensive experiments conducted on a multitype bridge data set demonstrate that BDDNet achieves high performance, attaining approximately 85% mean average precision (mAP) in damage detection and 82% mean intersection over union (mIoU) in segmentation. Visual results further confirm the model’s ability to accurately identify and delineate diverse damages, such as rust, missing bolts, cracks, and spalling, even in complex environments with limited training data. Overall, the proposed BDDNet provides an efficient, generalizable, and practical solution for UAV-based bridge inspection, enabling comprehensive structural condition assessment and laying the foundation for future quantitative damage evaluation.