An oriented object detection framework with enhanced feature extraction for real-time blade damage diagnosis in aero-engine borescope imagery

Abstract Borescope-based visual inspection is a crucial component of aero-engine health monitoring. However, its effectiveness is constrained by reliance on manual operations, which are labor-intensive, time-consuming and inefficient. To address these limitations, this paper proposes a deep learning approach named BDD-YOLO for automated damage detection and quantification on aero-engine blades. The model is built upon the YOLOv8 architecture and employs oriented bounding boxes (OBB) for precise detection. To enhance feature extraction, a novel C2F-CBAMDCNv2 module is introduced within the backbone network, which integrates deformable convolution and a convolutional block attention mechanism. Furthermore, the neck architecture incorporates improved bidirectional feature pyramid network to enhance multiscale feature fusion. Experimental results demonstrate that BDD-YOLO achieves a 2.3% increase in overall mean average precision (mAP) and a 5.1% increase in average precision (AP) specifically for the crack damage class. The model maintains an inference speed of 50FPS, enabling real time detection. For damage parameter quantification, the mean relative error is 3.74% for crack length measurement and 4.57% for ablation area measurement.