Mastering a lightweight method for the identification of pavement surface cracks plays a vital role in enhancing the efficiency and accuracy of pavement crack detection. However, the existing crack detection methods typically concentrate on improving model performance while neglecting the study of lightweight identification methods, causing inefficient crack detection. Therefore, this study aims to investigate a lightweight identification method for pavement cracks. High-definition vehicle-mounted cameras were employed to collect pavement crack images across multiple districts in Shenzhen, China, resulting in a dataset containing 20256 images. In addition, a pavement crack detection method based on the You Only Look Once version 8 (YOLO v8) model was proposed. On this basis, a knowledge distillation model enhanced by multiple teacher-assistants (KDMTA) was established to explore the impact of different teacher assistants and distillation paths on model performance. The results indicate that in the proposed KDMTA model, each level of teacher-assistant (TA) model can absorb diverse aspects of knowledge from the teacher models, thus broadening the range of insights. Subsequently, the information is imparted to lower-level TA models and the student model, expanding the perspective from which the student model acquires knowledge. Therefore, this improves the model's generalization performance and its effectiveness in target detection. The type and number of teacher assistant models, as well as the distillation path, influence the model performance. The random learning strategy enriches the combination and transfer of knowledge among teacher models, TA models, and student models, achieving lightweight identification for pavement crack. According to the results, the crack identification accuracy reaches 95.79 %, the mAP reaches 81.07 %, and the image processing speed has been improved by 79.6 %. This study enhances the efficiency and accuracy of pavement crack detection, providing methodological support for the rapid and accurate identification of pavement cracks. This study offers a novel method for the rapid and accurate identification of pavement cracks, which can assist management authorities to timely identify and repair pavement defects, reducing maintenance costs, and extending the lifespan of roads. Moreover, it can also facilitate the development of road safety management and intelligent transportation systems.
Read full abstract