| description abstract | Accurate recognition and location of pavement manholes are of great significance for pavement maintenance. This paper proposes an improved You only look once X (YOLOX) for automated detection of manholes on asphalt pavements. The proposed model improves the performance of the YOLOX model in two respects. First, the channel attention mechanism is introduced to enhance the model’s adaptive feature refinement; second, a microscale detection layer is deployed in the YOLOX model to extract more essential and distinct features. The experimental results are impressive, with the improved YOLOX achieving an F1 score and overall intersection-over-union of 98.14% and 91.61%, respectively, on 250 testing images, surpassing other state-of-the-art models such as YOLOv4, Faster R-CNN, EfficientDet, and the original YOLOX. To demonstrate robustness of the proposed model, the improved YOLOX is further applied to process manhole images taken randomly by a smartphone, which differ significantly from those acquired by a laser imaging system. It is found that the improved YOLOX can also yield similar detection efficiency in different scenes, which indicates the proposed model has a strong generalization ability. Particularly, the average frame per second (FPS) of the improved YOLOX is approximately 50.74 FPS using a modern graphic processing unit (GPU) device, implying the promising potential of the proposed model in supporting real-time automated detection of pavement manholes. | |
