Automated Detection of Pavement Manhole on Asphalt Pavements with an Improved YOLOX

• DOI: 10.1061/jitse4.iseng-2313
• Publication Date: 2023-07-26T12:57:07Z
• Authors: Hang Zhang; Jing Shang; Zishuo Dong; Anzheng He; Allen A. Zhang; Yang Liu; Kelvin C. P. Wang; Zhihao Lin
• Keywords: Robustness; Feature (linguistics)

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