Sub-millimetre level bridge crack measurement based on UAV and advanced MSDA-Net semantic segmentation

Abstract Traditional bridge crack detection methods based on manual inspection suffer from issues such as low efficiency, high cost, and poor safety. In recent years, the UAV and computer vision technologies have brought more efficient and secure solutions. However, due to flight safety, bridge crack images are captured from a distance. Existing deep learning networks cannot achieve sub-millimeter accuracy, thus failing to meet engineering requirements. Therefore, this paper proposes a multi-scale directional attention network (MSDA-Net) specifically targeting sub-millimeter level bridge crack measurement. In the encoder stage, the strip-SE module is proposed, explicitly enhancing the representation of directional features. In encoder–decoder fusion part, a multi-scale directional attention fusion module is constructed to capture local details and long-range dependencies while maintaining the continuity of the crack structure and the sharpness of the edges. The dual-weighted focal loss (DWFL) which can dynamically adjust the weight ratio between foreground and background to mitigate class imbalance is also designed. In addition, we proposed a measurement method based on geometric fitting and mathematical analysis, which elevates the measurement dimension from a discrete image space to a continuous mathematical geometric space. This method enables the calculation of dimensions with sub-millimeter accuracy. To validate the effectiveness of the method, the BRI_CRACK bridge crack dataset was constructed, and MSDA-Net was compared with eight SOTA models under identical experimental conditions. The results show that MSDA-Net performed best, with mean intersection over union (mIoU) and mean precision (mPrecision) improvements of 10.19% and 9.65% respectively over the next-best method. On the public dataset DeepCrack, the experiment results of MSDA-Net achieved mIoU, mPrecision, and mean recall of 90.82%, 95.74% and 96.23%, respectively. During field validation at the Kaijiang No. 5 Bridge in Deyang, Sichuan, the proposed method got a maximum error of only 0.1 mm in crack width and no more than 1% in crack length.