Performance of the four detection models.

A comparative study on automated pavement anomaly detection is conducted to improve the detection accuracy of models based on You Only Look Once version 4-Tiny (YOLOv4-Tiny). This study is the first to introduce a rotated rectangle labeling strategy for pavement anomaly detection. The pavement image dataset, primarily collected in Nantong, China, includes 1,107 cracks and 691 potholes. First, the YOLOv4-Tiny model is trained and validated as a baseline, achieving a mean average precision (mAP) below 0.4, which is inadequate for practical use. To improve performance, the YOLOv4-ResNet50 model is proposed by replacing the original backbone with a ResNet50 network, resulting in modest gains. To further enhance precision, rotated rectangles—bounding boxes that include an additional rotation angle—are used instead of conventional axis-aligned boxes. Accordingly, the YOLOv4-Tiny-rotated and YOLOv4-ResNet50-rotated models are developed and evaluated on the same dataset. Results show that the YOLOv4-ResNet50-rotated model achieves an mAP of 0.742, outperforming the more advanced YOLOX model, which reaches an mAP of 0.627. Moreover, the rotated bounding boxes enable more accurate representation of the shape and orientation of inclined cracks, making this model particularly well-suited for pavement anomaly detection. This study demonstrates the effectiveness of rotated rectangle labeling and rotated predicted bounding boxes in detecting inclined cracks and potholes, laying a foundation for further research in this area.