Real-time lightweight YOLO model for grouting defect detection in external post-tensioned ducts via infrared thermography

Abstract

It is challenging to distinguish the defective areas using infrared thermography to automatically analyze external post-tensioned tendon duct grouting defects. To achieve efficient and stable automated detection, a lightweight real-time grouting defects detection method based on YOLO deep learning is proposed. Firstly, the Cutpaste data augmentation method was used to effectively alleviate the problem of overfitting. Then, the C3Ghost module was introduced into the neck network, and the number of channels in the network layers was adjusted to 50 % of those in the YOLOv5s model, reducing the number of parameters and computational resources. Finally, the SGD optimizer and GIOU loss function, as well as the Sim attention module, were used to improve detection accuracy. Based on instance analysis and comparison, this method achieves mAP@0.5 of 96.9 % and detection speed of 66FPS. Compared with YOLOv5s, it reduces the number of parameters by 79 % and FLOPs by 77 %.