DBM-65k: A large-scale multi-scale dataset and benchmark for data-centric bridge damage identification

Abstract

UAV bridge damage inspection has long been limited by insufficient data scale, a single damage type, and lack of unified training paradigm for the model, making it difficult to achieve ideal results in real scenarios. This paper constructed the DBM-65k large-scale benchmark dataset and proposed the DBM (Detection Bridge Model) unified architecture model series to establish a comprehensive evaluation benchmark for bridge detection. Based on more than 65,000 images, DBM-65k constructed multi-scale and multi-category scenes, establishing a hierarchical perception system of "macro subject appearance and general micro-components". This paper uses YOLOv12 as the baseline, designs the dynamic alignment fusion module (DAF) and the dynamic detection target head (P2Head), and introduces adaptive scale-aware loss to amplify the characteristics of small-scale damage by dynamically adjusting the gradient return. For damage segmentation, a detection prior-driven cross-task transfer learning strategy is adopted to achieve collaborative training of detection and segmentation tasks. Experimental results show that the DBM series models show good performance in both macro-disease and micro-component tasks, especially achieving a significant gain of up to 5.37% in small object detection accuracy (mAPs). At the same time, DBM-seg provides high-precision pixel-level segmentation output. The training weights have been open sourced. This work will provide a solid data and algorithm basis for unified bridge structure detection.