Detection algorithm of aircraft skin defects based on improved YOLOv8n

In order to solve the problem of small targets being prone to false detection and missed detection in aircraft skin defect detection under complex backgrounds, the model of aircraft skin defect detection based on improved YOLOv8n is proposed in this paper. Firstly, the Shuffle Attention +  + module is incorporated into the network, combined with the residual connection idea, to more efficiently fuse feature map information; Secondly, SIOU and Focal Loss are used to replace CIOU as the regression loss functions to balance positive and negative samples in complex backgrounds and accelerate model convergence; Subsequently, the bidirectional feature pyramid network is used to modify the detection head and enhance multi-scale feature fusion. Furthermore, the depth-wise convolution module is used to replace the convolution module (Conv) in the neck part, which serves to reduce the parameters of the model and speed up the detection speed. Finally, an aircraft skin defect dataset is established, combined with Mosaic data enhancement to prevent the model from overfitting, and adopted the class balancing strategy to avoid class bias. The experimental results show that the detection accuracy of our improved YOLOv8n model is 97.9%, which is 7.3% higher than the baseline model. The model’s recall rate, the mean average precision, and F1 scores are improved by 13.9%, 6.6%, and 11.0%, respectively. The detection speed has achieved 139FPS, fulfilling the requirements of high accuracy and real-time performance in small target aircraft skin defect detection tasks.