Leveraging Data Efficient Image Transformer (DeIT) for Road Crack Detection and Classification
Humaira Anzum, Md. Nabil Sadd Sammo, Shamim Akhter
Abstract
Accurate detection and classification of road faults such as cracks is critical for transportation infrastructure maintenance. Road cracks impede comfortable traveling, endanger passenger safety, and create incidents. In this paper, we propose a new Data Efficient Image Transformer (DeIT)-based Road crack identification system that can distinguish between cracks and non-cracks in road surfaces. This study also evaluates how well DeIT performs in comparison with YOLOv5, YOLOv8, Xception, ResNet50 and MobileNetV2 deep learning algorithms. We have used a publicly available dataset consisting of road images from Kaggle Concrete Data to train and test our model. For identifying cracked and non-cracked road surfaces, our proposed model achieved the highest 99.75 % accuracy on the test dataset using DeIT algorithm.