A lightweight L-YOLO-based method for small-scale damage detection in aero-engines

To address the issues of poor real-time performance and low detection accuracy when deploying deep learning models for aero-engine damage detection on embedded devices, a Lightweight-You Only Look Once (L-YOLO) model for small-scale damage detection in aero-engines is proposed. Firstly, the C2f structure of the YOLOv8 backbone network is reconstructed using a Lightweight Multi-Scale Fusion Convolution (LMSFConv), enhancing the model's multi-scale feature extraction capabilities while reducing computational load. Secondly, the YOLOv8 neck network is improved using an Lightweight Multi-Scale Feature Fusion Network (LMSFFN), allowing the model to integrate more small-scale damage information while reducing the complexity of the neck network. Finally, experimental validation is conducted on an aero-engine damage dataset. The results indicate that, compared to the original model, the L-YOLO model reduces the number of parameters by 34.5% and the number of Giga Floating-point Operations (GFLOPs) by 10.2%. The Average Precision (AP) for Burn, Crack, Dent, and Material Missing damages improves by 0.6%, 0.6%, 3.1%, and 1.1%, respectively, with an overall mAP improvement of 1.4%. On embedded devices, the mAP reaches 95.1%, higher than other models, and the Frames Per Second (FPS) reaches 62, matching the image acquisition speed for engine damage, thus better meeting the intelligent application needs of aero-engine damage detection.