Road damage detection using transformers an‎d semi-supervised learning

10/20/2026 12:00:00 AM

Road damage detection is a vital process for identifying an‎d assessing existing defects an‎d damages on road surfaces, which plays an important role in improving transportation safety, reducing maintenance costs, an‎d preventing road accidents. Given the expansion of road networks an‎d the need for continuous monitoring of infrastructure conditions, the use of intelligent an‎d machine vision-based methods for automatic detection of road damage, especially in urban environments, has become one of the key challenges in the field of artificial intelligence an‎d image processing. In this study, in order to provide specific an‎d real data, a database (dataset) of images of urban roads in Jouybar County was collected. In order to increase the volume an‎d diversity of data an‎d improve the power of learning models, a set of data augmentation techniques were applied to the images. Subsequently, three versions of powerful YOLO models, including YOLOv5, YOLOv8, an‎d YOLOv10, were trained on the prepared data. The eva‎luation results of these models showed that the models were able to achieve accuracies of 0.858, 0.696, an‎d 0.882 in the mAP@0.5 criterion, respectively, which indicates the high ability of the YOLOv10 model in accurately detecting road damage. In the next step, with the aim of utilizing unlabeled data an‎d reducing the need for manually labeled data, the semi-supervised learning method was used. The results obtained from this step showed that combining labeled an‎d unlabeled data improved the performance of the model an‎d increased its accuracy in the mAP@0.5 criterion to 0.799. Finally, to achieve an optimal model with higher performance, a new proposed architecture called C3TR was used, which is a combination of the capabilities of the YOLOv5 model an‎d the transformer structure. This model, by utilizing the attention mechanism in the transformer structure, was able to provide significant performance in road damage detection an‎d achieved an accuracy of 0.858 in the mAP@0.5 criterion. The results of this research show that combining advanced deep learning structures, especially attention-based models, with object recognition architectures such as YOLO can significantly increase the efficiency of road damage detection systems an‎d be an effective step towards developing smart urban road maintenance solutions.

تشخيص آسيب جاده , ماشين لرنينگ , ترنسفورمر , يادگيري نيمه نظارتي

Road damage detection , machine learning , Transformer , Semi-supervised learning

Masoumeh Razi Pour Juybari

Dr. Abdollah Amirkhani