Deep learning for automated multiclass surface damage detection in bridge inspections

This paper presents a deep learning-based approach for multiclass surface damage detection and segmentation in various bridge components. The proposed BridgeNet integrates advanced techniques including the Swin Transformer, the CARAFE upsampler, and transfer learning to enhance damage identification. Furthermore, a comprehensive dataset, named BridgeDamage is established, which consists of over 2800 annotated bridge inspection images, covering five major categories of surface defects. Experimental validations of BridgeNet demonstrate the effectiveness and robustness of the proposed approach, allowing the distinct detection and clear segmentation of various types and instances of bridge damage. In the comparative experiment, BridgeNet exhibits a substantial improvement in both mAP and mIoU metrics, surpassing the original Mask R-CNN by more than 33% and 26%, respectively, and outperforming other state-of-the-art deep learning models, with a maximum mAP of 74.7% and mIoU of 66%. The results highlight the promising potential of the proposed approach for practical applications in bridge inspection.