Surface Defect Detection of Aircraft Glass Canopy Based on Improved YOLOv4

Abstract

Aiming at the problems of low detection accuracy and high false detection rate in traditional defect detection methods, an improved YOLOv4 model for surface defect detection of aircraft glass canopy is proposed. In this paper, an improved FReLU activation function is used instead of Mish to better adaptively capture the spatial correlation to improve the defect detection efficiency. It is worth mentioning that due to the high aspect ratio of defects, this paper defines the rotated rectangle method without adding rotation anchors, which greatly improves the prediction speed of the model. Finally, this paper solves the problem of model fitting by optimizing the loss function of the model. Experiments show that the mAP value of our model on the test set reaches 93.34 %, which is 4.58 % higher than original YOLOv4 model, which proves the effectiveness of our model on the aircraft glass canopy surface defect dataset.