Road Disease Damage Detection Based on Improved CenterNet

Road damage detection is an important step in ensuring road safety and achieving timely repair of road damage, which is of great significance in reducing road maintenance costs. Deep learning technology has developed rapidly this year, and using computer vision technology instead of manual detection can achieve automated detection of road losses. However, there are still a series of problems such as low detection efficiency and poor image quality. In response to the above issues, in this article, we first established a road dataset, including 2575 Chinese road images and 3589 road damage annotations. Then, we proposed an improved model based on CenterNet, using the Asymptotic Feature Pyramid structure. Using this method, we fused two adjacent low-level features, and gradually fused the high-level features. In the feature fusion process, adaptive spatial fusion is utilized to avoid multi-target information conflict, and then the CLA cross-layer attention mechanism is utilized to improve the performance of small-target detection in the model and reduce the computational amount of the model. Finally, it is experimentally demonstrated that our improvement improves the detection accuracy and the computational cost is within a manageable range.