Pixel-Level patch detection from full-scale asphalt pavement images based on deep learning

The patch recognition and localisation from high-resolution pavement images play a key role in evaluating asphalt pavement condition. To achieve the purpose, two line-scan cameras were employed to acquire lots of full-scale asphalt pavement images with 1 mm² per pixel resolution in the field. The raw images were pre-processed to construct the dataset, including 984 patches of various shapes in 827 pavement images. Subsequently, the YOLO series models innovatively applied in automatically detecting the patches from pavement images were well trained with the pavement training and validation sets. The testing results reveal that the F1-score and mAP@0.5 values of the YOLOv4 model are 0.911 and 92.92% respectively in the test set. The comprehensive recognition accuracy of the YOLOv4 model outperforms the other YOLO models. And the detection speed of the YOLOv4 model for pavement data video is 38.1 frames per second. Additionally, all the features of patches in the partial testing pavement images, no matter whether they are strip, block, intersection, blurry within the shadow, or background noise, can be correctly predictedby the rectangular box. The average localisation errors of the width and height of the predicted patches are 21 and 16 mm respectively, and the average IoU is 0.83.