TITLE:
SCD-YOLO: A Detail-Enhanced Network for Strip Steel Surface Defect Detection
AUTHORS:
Jiacheng Jiang, Yuebin Su
KEYWORDS:
Strip Steel Surface Defect Detection, YOLOv11n, Star-Context Enhanced C3k2, Detail-Enhanced Detection Head
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.7,
July
30,
2026
ABSTRACT: Accurate and real-time surface defect detection is essential for industrial visual inspection. However, strip steel defects often show weak textures, ambiguous boundaries, low contrast, and strong background interference. These characteristics make fine-grained defect representation and accurate localization challenging. Although recent YOLO-based detectors achieve high inference efficiency, their convolution-dominated feature extraction modules and generic prediction heads remain limited in modeling subtle defect cues. To address these issues, this paper proposes SCD-YOLO, a detail-aware real-time detector built upon YOLOv11n. Extending our previous Star-based feature interaction and lightweight shared detection head, we introduce two defect-aware enhancements. First, a Star-Context Enhanced C3k2 structure, termed SC-C3k2, is developed for feature extraction. It embeds context anchor attention into the star-shaped multiplicative interaction path. This design performs contextual recalibration on high-order nonlinear interaction features. As a result, it strengthens weak-texture defect representation and suppresses pseudo-defect activations from complex backgrounds. Second, a Lightweight Shared Detail-Enhanced Convolutional Detection Head, termed LSDECD, is proposed for prediction. It incorporates detail-enhanced convolution into the shared multi-scale prediction transformation. This design reduces prediction-head redundancy and improves sensitivity to local edges, texture discontinuities, and subtle intensity variations. Experiments on the NEU-DET dataset show that SCD-YOLO improves the [email protected] from 75.64% to 78.81% compared with YOLOv11n. The absolute gain is 3.17 percentage points, while the model maintains real-time inference speed. Ablation studies further confirm that SC-C3k2 and LSDECD provide complementary improvements in feature discrimination and detail-aware localization. These results demonstrate that SCD-YOLO achieves a favorable accuracy-efficiency trade-off for real-time strip steel surface defect detection.