FSU-YOLO: aero-engine blade defect detection based on improved YOLOv8

Abstract Defect inspection of aero-engine blades represents a critical approach to assessing blade integrity and ensuring aircraft flight safety; however, artificial vision-based inspection relies heavily on operator expertise and is therefore susceptible to human factors. Therefore, this paper proposes a lightweight detection network, FSU-YOLO, based on YOLOv8. To achieve better adaptive frequency detail enhancement, a design proposal for the frequency-enhanced efficient module is presented, which enhances the model’s ability to perceive complex defect features while maintaining computational efficiency. On this basis, to highlight the temporal correlation of data collected by the borescope equipment, the spatio-temporal event-driven attention bridge module is designed to capture salient spatial-temporal relationships and improve the detection precision of minor defects. To improve the detection precision of targets at different scales and enhance the response to critical defect regions, a UCB-based lightweight fusion block is designed for spatial feature extraction, thereby increasing the attention of the detection network to defects. The experimental results show that the mAP@50 of FSU-YOLO in the constructed aero-engine blade surface defects dataset reaches 91.1%, which is a 2.7% enhancement compared with the base model; in the NEU-DET dataset, the mAP@50 is enhanced by 3.5%. When deployed on a mobile device with an AMD Ryzen 7 7840U, FSU-YOLO achieves 28.25 frames per second with 35.40 ms of latency, meeting real-time industrial inspection requirements. The experimental results verify the feasibility and generalizability of the proposed method, demonstrating that FSU-YOLO can be effectively applied to the task of surface defect detection.