计算机科学
稳健性(进化)
串联(数学)
人工智能
残余物
融合
探测器
推论
模式识别(心理学)
材料科学
特征提取
计算机视觉
可靠性(半导体)
分割
表面贴装技术
特征(语言学)
烧蚀
灵敏度(控制系统)
内存占用
测距
绩效改进
渲染(计算机图形)
低分辨率
作者
Weili Chen,Jianjing Wei,Xin Guo,Bidong Chen
标识
DOI:10.1038/s41598-026-64588-4
摘要
Steel surface defect detection remains challenging due to low-contrast textures, fine crack-like patterns, and weakly textured granular defects, particularly under uneven illumination. This paper presents SFEDet, an improved detector built upon YOLO12 with three tailored designs. First, Learnable Gray Preparation (LGP) performs learnable contrast enhancement at the input stage via a gated residual formulation. Second, DefectFreqFusion replaces conventional concatenation with spatial–frequency dual-path fusion for robust multi-scale feature integration. Third, an edge- and morphology-aware neck enhancement scheme is developed, combining C2fDE for multi-scale receptive fields, Edge-Guided Attention Module (EGAM) for Sobel edge-guided attention, and Efficient Head Preparation (EfficientHeadPre) for lightweight pre-detection refinement; together these modules constitute the Bottom-Up Steel Neck (BU-SN). Experiments on NEU-DET (640$$\times$$640) demonstrate that SFEDet-Large achieves 75.1% mAP@0.5, outperforming the YOLO12-L baseline by 2.5 percentage points, while the lightweight SFEDet-Nano variant attains 72.6% mAP@0.5 with only 8.65M parameters and 333 FPS inference speed. Ablation studies confirm BU-SN as the primary contributor to detection accuracy, while LGP significantly enhances robustness under varying illumination conditions, yielding an 8.5% mAP improvement in extremely dark scenarios. The proposed framework offers favorable accuracy–complexity trade-offs for industrial inspection scenarios, with SFEDet-Large prioritizing detection reliability and SFEDet-Nano enabling efficient real-time deployment.
科研通智能强力驱动
Strongly Powered by AbleSci AI