Research on Defect Rotation Detection of Aircraft Flared Tube Based on Improved YOLOv4

Abstract

Aircraft Flared Tube (AFT) is an important part of the aircraft system, and its surface defect detection has become a prerequisite for meeting the long-term normal operation of the aircraft. Although the existing deep learning defect detection methods have made great progress, there are still problems such as difficulty in detecting tiny defect, insufficient model generalization ability and harsh detection environment. Therefore, based on the analysis of the YOLOv4 model, we first utilize the Multistage Attention Module (MAM) to enhance the feature expression ability of the shallow network. Secondly, we exploit rotation detection to accommodate large aspect ratio defects on AFT surfaces. Finally, we convert the pytorch model into a tensorRT engine and deploy it on Agx xavier to achieve high efficiency, high-stability, and low-power inference. Experimental results show that the mean Average Precision (MAP) of our improved model reaches 97.87%, and the single image detection speed reaches 223.23ms, which further proves the good performance of our model on the AFT detection task.