An Enhanced Aircraft Fuselage Defect Detection Model With Hybrid Logical-Feature Distillation

Small defects result in the loss of essential information during defect detection, and enhancing the detection accuracy of these minor defects is a widely researched direction, especially in the development of detection network that achieve high accuracy while simultaneously maintaining low model complexity. To strike a balance between accuracy and efficiency, this paper introduces a high-efficiency aircraft fuselage defect detection model. Firstly, we restructured a teacher model consisting of parallel backbone with Aux_IoU. This redesign enhances the detection of small defects by improving the characterization capability of the feature extraction network and implementing fine-grained bounding boxes. Secondly, to decrease the complexity of teacher model while preserving high accuracy, we introduce a hybrid logical-feature distillation framework. The student model is trained to assimilate the teacher’s feature information and logits through mask generative distillation (MGD) and logical distillation, respectively. Finally, to validate the effectiveness of proposed network, we conducted experiments on Aircraft_Fuselage_DET2023 dataset. The experimental results reveal that our student model achieves enhancements of 8.3% and 2.7% over the baseline YOLOv8n on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mAP</i><sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</sub> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mAP</i><sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50:95</sub>, respectively. Furthermore, it demonstrates improvements of 0.5% and 0.2% compared to the teacher model, while simultaneously reducing the number of parameters and computational complexity by 47.3% and 53.9%. In comparison to mainstream object detection algorithms, our model achieved superior performance.