Asphalt Crack Detection and Segmentation Using Deep Learning

Abstract

Semantic segmentation is a computer vision task that utilizes deep learning algorithms to recognize a collection of pixels that form a distinct class. This technique could be used to early recognize road pavement cracks and reduce maintinence cost and inssuring safety for all road users. This research study presents an interesting semantic segmentation model for detecting asphalt cracks in roads based on deep learning techniques that combine object detection and semantic segmentation through three steps: In the first step, preprocessing the source images, then YOLOv10 model had been used for the crack detection framework. Finally, the UNet 3+ model was employed as a semantic segmentation model in which pixel-level segmentation is carried out. The geometric properties of the cracks are quantified to assess the damage in the road. The system has been trained, evaluated, and tested using two datasets: The SUT-Crack dataset and the IRD-Crack dataset. The proposed system shows excellent performance across different metrics such as Recall, Precision, Accuracy, mAP, Confidence score, and Dice coefficient. The accuracy reached up to 99.06%, demonstrating its ability to be applied in real- world environments.