Hybrid Deep Learning Framework for Crack Detection and Condition Classification of Prestressed Concrete Sleepers

Abstract

Railway sleeper degradation is a major problem for track safety and maintenance planning, especially in real–world scenarios where ballast occlusion, variable lighting, and weather–induced surface changes make visual inspection difficult. Surface–based image analysis can also be affected by complex backgrounds, while conventional manual inspection is labour–intensive and subjective. To overcome these issues, this study presents a hybrid deep learning framework that combines image–level sleeper condition classification with object–level sleeper and crack detection using field images collected from the LFI Arezzo corridor, Italy, including images of in–service and discarded sleepers. The classification dataset comprised 294 images and was evaluated using EfficientNet–B0, ResNet50 and VGG19 under stratified five–fold cross–validation. EfficientNet–B0 and ResNet50 achieved 100% mean accuracy, while VGG19 achieved 99.66%. Grad–CAM was used to examine the spatial regions contributing to the classification decisions. For object detection, YOLOv11, RT–DETR and Faster R–CNN were evaluated on a common validation set. RT–DETR obtained the highest mAP@0.5 of 0.740, YOLOv11 recorded the highest overall precision of 0.769, and Faster R–CNN achieved the highest F1–score of 0.681. Sleeper localization was consistently strong across all three detectors, while fine and low–contrast cracks remained the more difficult class. The results show that the three detector families offer complementary strengths for field–based sleeper inspection and provide a practical basis for AI–assisted screening, localization and crack review in railway maintenance workflows.