Automated Pavement Crack Detection with Deep Learning Methods: What Are the Main Factors and How to Improve the Performance?

Recently, deep learning (DL) methods have been intensively studied in pavement distress detection research. To build an effective yet robust DL-based distress detection model, both the training data and DL model structure need to be carefully prepared. This paper discusses the main influencing factors that affect the performance of DL models in pavement distress detection and aims to explore possible ways to improve the model performance. An open-source pavement image dataset and three promising object detection methods were selected for model development and evaluation. A series of experiments were conducted with different settings of model structures and data manipulation. According to the experiment results, selection of detection methods, data size, and annotation consistency have significant impact on detection performance. DL models YOLO and Faster R-CNN were found to yield the same level of performance, while CenterNet notably underperformed on the test dataset. Distress classes with a larger number of instances were found to be able to yield better performances. With the new annotation data carefully marked up and verified by experienced pavement raters, the accuracy score using the F1 index was found to be 0.84, which was a very sound improvement compared with the 0.7 of the original annotation. Data augmentation methods were found to be able to significantly improve the detection accuracy. Based on the experiment results, we believe the focus should be placed on the development of high-quality training datasets and thin object detection methods in the future.