Advancing UAV-based Inspection System: The USSA-Net Segmentation Approach to Crack Quantification

In the realm of crack inspection for complex infrastructures, traditional methods have primarily relied on expensive structural health monitoring instruments and labor-intensive procedures. The emergence of unmanned aerial vehicle (UAV) technology brings about effective and innovative solutions for bridge inspection. To advance the technology, this study presents a novel crack inspection system that employs light detection and ranging (LiDAR) scanning to construct a 3-D model of the target structure. A path planner is then developed to ensure complete coverage of all crack points on the structure being inspected. Through extensive testing, the proposed system demonstrates successful detection and localization of various types of cracks. Furthermore, our improved deep crack segmentation model, U-Net with spectral block and self-attention module, surpasses the performance of the original U-Net model, exhibiting a 3.2% higher Dice coefficient and a 3.3% higher mean intersection over union (mIoU) evaluation metric on our self-established crack dataset. In the case of the Crack500 public dataset, our model outperforms the original U-Net model by 10% in Dice coefficient and 14% in mIoU. Moreover, our U-Net with spectral block and self-attention module (USSA-Net) outperforms other latest state-of-the-art (SOTA) models on the DeepCrack500 dataset, surpassing the progressive and adaptive fusion (PAF)-Net and progressive and hierarchical context fusion (PHCF)-Net by approximately 5% in Dice coefficient and 2.7% in mIoU. For crack size estimations, our proposed system accurately estimates the horizontal and vertical dimensions of cracks, achieving a root-mean-square error (RMSE) of 9.9 and 6.2 mm, respectively. Overall, the system achieves millimeter-level crack size estimation accuracy. Moreover, our system is characterized by its low-cost nature and lightweight design. Experimental results showcase the system's robustness and effectiveness in executing real-world crack inspection tasks, even within complex environments.