CCP-YOLO: An Improved YOLOv11n Algorithm for Steel Surface Defect Detection

Abstract

Highlights What are the main findings? CCP-YOLO improves YOLOv11n for steel surface defect detection by enhancing multi-scale feature extraction, heterogeneous feature fusion, spatial-detail-preserving context aggregation, and bounding box regression. The proposed model achieves 80.2% mAP50 on the NEU-DET dataset, improving the YOLOv11n baseline by 4.1 percentage points, while maintaining 2.7 M parameters, 6.6 GFLOPs, and an inference speed of 133.14 FPS. What is the implication of the main findings? CCP-YOLO provides an accurate and lightweight solution for detecting small, multi-scale, and irregular steel surface defects in industrial inspection scenarios, and its improved performance on both NEU-DET and GC10-DET demonstrates strong generalization potential for practical steel surface quality inspection systems. Abstract Steel surface defect detection remains challenging due to difficulties in multi-scale feature extraction, limited effectiveness of heterogeneous feature fusion, and loss of spatial detail information. To address these issues, this paper proposes CCP-YOLO, an improved steel surface defect detection algorithm based on YOLOv11n. The proposed method introduces four targeted improvements: (1) a Multi-Scale Dilated Reparameterization module (C3k2_MSD) for enhanced multi-scale feature extraction via a three-branch parallel reparameterization architecture; (2) an Interactive Adaptive Feature Fusion Module (IAFM) for effective integration of heterogeneous features; (3) a Progressive Shared-Weight Context Aggregation (PSWCA) module replacing the original SPPF structure to preserve spatial detail; and (4) a Wise-Inner-MPDIoU fusion loss function for improved bounding box regression accuracy and stability. Experimental results on the NEU-DET dataset demonstrate that CCP-YOLO achieves an mAP50 of 80.2%, representing a 4.1 percentage-point improvement over the YOLOv11n baseline, with a recall of 0.754, 2.7 M parameters, 6.6 GFLOPs, and an inference speed of 133.14 FPS. Further validation on the GC10-DET dataset confirms a 3.7 percentage-point improvement in mAP50. These results indicate that CCP-YOLO effectively enhances detection accuracy while maintaining computational efficiency, demonstrating strong potential for real-world industrial deployment.