Speaker
Description
Recent advancements in Transformer-based deep learning have opened new avenues for precise and robust defect detection in critical transportation infrastructure components, such as tunnels, rails, and pavements. This paper unifies three related investigations on Transformer-driven detection frameworks. First, by enhancing Swin Transformer for tunnel lining crack inspection, crack features were more accurately segmented under challenging illumination and structural conditions, leading to a notable increase in mean Average Precision (mAP) and detection speed. Second, the proposed “RailFormer” introduced a Transformer-based encoder-decoder architecture for Rail Surface Defect (RSD) detection, employing both global and local feature fusion. This design outperformed traditional CNN approaches by capturing fine-grained hierarchical linkages in rail surface imagery, thus achieving higher mean Intersection over Union (mIoU). Lastly, a dedicated Transformer-based semantic segmentation approach was devised for pavement crack detection, integrating the Swin Transformer as the Encoder with an attention-fused Decoder, significantly boosting both the mean F1 (mF1) and mean Recall (mRecall) scores over other semantic segmentation models. Overall, these studies underscore the Transformer’s capacity to learn long-range dependencies and detail-rich features, overcoming the inherent limitations of CNN-based methods when dealing with thin, noisy, and complex cracks. By drawing on the key innovations and demonstrated performance gains in different infrastructure scenarios, this research highlights the versatility and effectiveness of Transformer architectures for transportation infrastructure defect detection, offering a promising path toward more accurate, efficient, and automated maintenance workflows.
