CAD-YOLO11: a lightweight class-aware dynamic optimization framework for pipeline defect detection

Pipeline defect detection is crucial for ensuring the safety and reliability of industrial infrastructure. However, lightweight object detectors often suffer from insufficient feature representation and category-specific optimization capability when dealing with small-scale defects, weak textures, and visually similar defect categories. This paper proposes CAD-YOLO11, a lightweight class-aware dynamic optimization framework based on YOLO11 for accurate and efficient pipeline defect detection. The framework incorporates an adaptive downsampling strategy to improve fine-grained feature preservation during feature compression, while introducing a DetectAux_ClsEnhance branch to provide additional category-level supervision for enhancing semantic discrimination. Furthermore, a Class-aware Weighted Dynamic Optimization (CWDO) strategy is developed, including Class-Aware Dynamic Top-k Assignment (CADT) and Class-Aware Coordinate Optimization (CA-CN), to achieve category-adaptive positive sample selection and bounding box optimization according to different defect characteristics.Experiments conducted on the Daji pipeline defect dataset demonstrate the effectiveness of CAD-YOLO11. Compared with the original YOLO11n, the proposed method improves mAP50 from 88.64% to 91.13% and mAP50-95 from 66.52% to 69.90%, while reducing computational complexity from 6.3 GFLOPs to 5.3 GFLOPs. The proposed framework achieves a favorable balance between detection accuracy and computational efficiency, providing an effective solution for real-time industrial pipeline defect inspection.