Scholarlink

<p>Real-time detection of small surface defects in high-speed hot-rolling production lines remains challenging due to strong rolling textures, low-contrast backgrounds, and strict latency constraints. To break this bottleneck between accuracy and latency, this paper presents SF-YOLO11, a lightweight spatial-frequency collaborative network built upon the Ultralytics YOLO11 framework. This work addresses the spectral bias of convolutional neural networks toward low-frequency components by introducing a dual-domain mechanism that enhances high-frequency micro-defect features in a computationally efficient manner. Specifically, a multi-scale Gaussian edge attention (MGEA) module is deployed in the shallow layers. By leveraging fixed multi-scale Gaussian kernels and Scharr operators, MGEA explicitly decouples defect boundaries from texture aliasing and strengthens geometric edge cues with negligible additional parameters. Furthermore, a spatial-frequency fusion attention module is designed, which integrates a learnable soft high-pass mask in the discrete cosine transform domain with a spatial large-kernel depthwise separable convolution branch. This design selectively amplifies defect-related high-frequency responses while suppressing background noise through complementary spatial-frequency perception. Experimental results on NEU-DET, GC10-DET, and a proprietary Steel-Defect dataset demonstrate that SF-YOLO11 achieves an excellent speed–accuracy trade-off; compared with the YOLO11 baseline, it improves mAP@0.5 on NEU-DET by 4.2% (up to 81.1%) with only an additional 0.3M parameters and 1.6 GFLOPs. Running at 275 FPS on an NVIDIA RTX&#xa0;4090 GPU, the proposed method substantially reduces both false negatives and false positives, and satisfies the stringent real-time requirements of modern industrial quality-control systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SF-YOLO11: a spatial-frequency collaborative network for real-time steel surface defect detection

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Tianyu Liu,
  • Haiyu Li,
  • Shilong Wu,
  • Haoran Dong

摘要

Real-time detection of small surface defects in high-speed hot-rolling production lines remains challenging due to strong rolling textures, low-contrast backgrounds, and strict latency constraints. To break this bottleneck between accuracy and latency, this paper presents SF-YOLO11, a lightweight spatial-frequency collaborative network built upon the Ultralytics YOLO11 framework. This work addresses the spectral bias of convolutional neural networks toward low-frequency components by introducing a dual-domain mechanism that enhances high-frequency micro-defect features in a computationally efficient manner. Specifically, a multi-scale Gaussian edge attention (MGEA) module is deployed in the shallow layers. By leveraging fixed multi-scale Gaussian kernels and Scharr operators, MGEA explicitly decouples defect boundaries from texture aliasing and strengthens geometric edge cues with negligible additional parameters. Furthermore, a spatial-frequency fusion attention module is designed, which integrates a learnable soft high-pass mask in the discrete cosine transform domain with a spatial large-kernel depthwise separable convolution branch. This design selectively amplifies defect-related high-frequency responses while suppressing background noise through complementary spatial-frequency perception. Experimental results on NEU-DET, GC10-DET, and a proprietary Steel-Defect dataset demonstrate that SF-YOLO11 achieves an excellent speed–accuracy trade-off; compared with the YOLO11 baseline, it improves mAP@0.5 on NEU-DET by 4.2% (up to 81.1%) with only an additional 0.3M parameters and 1.6 GFLOPs. Running at 275 FPS on an NVIDIA RTX 4090 GPU, the proposed method substantially reduces both false negatives and false positives, and satisfies the stringent real-time requirements of modern industrial quality-control systems.