A real-time multiscale pavement crack-detection model A real-time multiscale pavement crack-detection model

• Published in: Signal, image and video processing Vol. 19; no. 14; p. 1237
• Main Authors: Wang, Xingju, Liang, Yiqian, Yang, Xingyu, Guo, Mengqi, Jia, Tongyao, Kamran, Muhammad
• Format: Journal Article
• Language: English
• Published: London Springer London 01.12.2025; Springer Nature B.V
• Subjects: Accuracy; Computer Imaging; Computer Science; Cracks; Feature extraction; Image Processing and Computer Vision; Multimedia Information Systems; Original Paper; Pattern Recognition and Graphics; Real time; Semantics; Service life; Signal,Image and Speech Processing; Unmanned aerial vehicles; Vision; Transformer-based object detection; Multiscale feature fusion; Real-time deep learning; Pavement crack detection
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-025-04836-8

Intelligent detection of pavement cracks is crucial for infrastructure maintenance, where timely detection of cracks and damage significantly prolongs pavement service life. However, current detection models still have limitations in terms of accuracy. To address this shortcoming, this study proposes a pavement crack real-time detection transformer (PC-DETR) for accurate real-time pavement crack detection. PC-DETR, based on the real-time detection transformer (RT-DETR), integrates the SENetV2 feature fusion mechanism, semantic and detail infusion mechanism, and inner-Focaler-weighted intersection-over-union(IFWIoU) loss function. PC-DETR, the first method to use the RT-DETR model in the field of pavement crack detection, determines the proposed IFWIoU loss function, which combines weighted intersection-over-union (IoU), Focaler-IoU, and Inner-IoU functions. This integration strengthens global feature extraction, reduces information loss in minor defects, and enhances the synergy between the deep and shallow feature layers. Our experimental validation on the RDD2022 dataset demonstrated the superior performance of the proposed model: PC-DETR achieved 55.3 GFLOPs. A road surface inspection vehicle using this method was found to effectively detect pavement cracks in real scenarios at a speed of 40 frames per second.