CAM-RTDETR: an aerospace engine surface defect detection method based on improved transformer architecture

To address challenges in the surface defect detection of core Aerospace Engine components, specifically targeting texture aliasing under strong metallic backgrounds and the difficulty in identifying cross-scale minute defects such as tiny dots, this paper proposes CAM-RTDETR, a detection model based on an improved RT-DETR-r18. First, to accommodate the irregular geometry of blades, a C2f-DBlock backbone based on Multi-Branch Dilated Convolutions is designed to effectively decouple defect features from the rigid background grid. Second, a Texture-Sensitive Self-Attention (AIFI-TSSA) module is introduced into the Encoder, incorporating learnable texture biases to suppress the interference of high-frequency metallic grain noise. Finally, a Mixed Aggregation Network (MANet) is adopted to replace the standard fusion module, utilizing a multi-branch structure to preserve the faint signals of minute defects during feature fusion. Experimental results on a self-constructed Aerospace Engine defect dataset indicate that, while maintaining real-time performance, the proposed method reduces parameter count by 21.5% and computational cost (GFLOPs) by 16.1% compared to the baseline RT-DETR-r18. Furthermore, it achieves a 1.7% increase in mAP@0.5–0.95 and a 4.0% increase in Precision. Compared with mainstream YOLO series models, the method demonstrates significant performance advantages in detecting minute defects and low-contrast scratches.