CGDD-YOLO: a recognition and classification algorithm for small parts based on YOLO11

To address the detection challenges of multi-category defect-free parts in industrial non-destructive testing (NDT), this paper proposes CGDD-YOLO, an improved algorithm based on the YOLO11 framework. Drawing on the DRB module concept, we integrate the DWR module from DWRSeg into the C3k2 module, leveraging large-kernel convolutions and depth-wise separable dilated convolutions to enhance efficient multi-scale feature extraction. Inspired by CGNet, a CGDown down sampling module with an attention mechanism replaces the original convolutional layer to fuse local and global contextual information, while the DySample module substitutes traditional up sampling for dynamic and efficient feature restoration. Additionally, the YOLO11 detection head is enhanced with Dynamic Head, which integrates scale, spatial, and task attention to boost detection performance. Experimental results on a multi-part type dataset demonstrate that CGDD-YOLO outperforms YOLO11n: mAP@0.5 reaches 99.2% (a 0.2% increase) and mAP@0.5:0.95 achieves 90.9% (a 3.7% improvement). This work effectively improves the detection accuracy of the network, providing a reliable algorithmic solution for the non-destructive testing of various components in factory environments.