| 黄起豹,周在峰,李治国,胡志永.基于重参数化注意力轻量化YOLOv8的纸张缺陷高效检测方法研究[J].中国造纸,2026,45(6):222-229 |
二维码（扫一下试试看！） |
| 基于重参数化注意力轻量化YOLOv8的纸张缺陷高效检测方法研究 |
| Study on an Efficient Detection Method Based on Lightweight YOLOv8 with Reparameterized Attention |
| 收稿日期：2026-02-02 修订日期：2026-02-22 |
| DOI：10.11980/j.issn.0254-508X.2026.06.025 |
| 关键词: 纸张缺陷检测 在线质量控制 YOLOv8模型 重参数化卷积 轻量化神经网络 |
| Key Words:paper defect detection online quality control YOLOv8 model re-parameterized convolution lightweight neural network |
| 基金项目:中国索引学会人工智能研究项目基金（CSI25C01）；中国成人教育协会人工智能一般项目（AI-Y2025028S）；中国机械工业教育协会（ZJJX25SY001）；全国高等学校计算机教育研究会（CERACU2026RO7）。 |
|
| 摘要点击次数: 197 |
| 全文下载次数: 15 |
| 摘要:纸张缺陷在线检测精度与实时性的协同优化，是高速造纸生产线质量控制的核心瓶颈。本研究以YOLOv8n为基准模型，提出融合重参数化卷积与位置敏感多查询注意力（C2PMQA）的轻量化框架，通过部分深度可分离卷积（PDSConv）、重参数化ConvNeXt增强型C2f（RCNXC2f）、C2PMQA三大核心模块协同优化，实现参数精简、多尺度特征提取与全局-局部特征融合，并采用结合轻量化检测头（LWHead）的非极大值抑制（NMS-Free）策略实现端到端推理，构建了重参数化注意力轻量化YOLOv8模型。基于黑斑、孔洞、褶皱、划痕4类典型缺陷数据集的验证表明，该模型的特征提取能力、参数效率与推理速度均有所提升，平均精度均值（mAP₅₀）达99.2%，参数量仅1.35 MB，计算量3.07 GFLOPs，在RTX 4060 GPU上推理帧率达175 FPS，可适配600 m/min高速纸机需求。 |
| Abstract:The synergistic optimization of accuracy and real-time performance in online paper defect detection represents a core bottleneck for quality control in high-speed papermaking production lines. Taking YOLOv8n as the baseline model， the study proposed a lightweight framework that integrated re-parameterized convolution with position-sensitive multi-query attention (C2PMQA). Through the collaborative optimization of three core modules： partial depthwise separable convolution (PDSConv)， re-parameterized ConvNeXt-enhanced C2f (RCNXC2f)， and C2PMQA， the proposed framework achieved parameter reduction， multi-scale feature extraction， and global-local feature fusion. Furthermore， a non-maximum suppression (NMS-Free) strategy incorporating a lightweight detection head (LWHead) was adopted to enable end-to-end inference， resulting in a re-parameterized attention lightweight YOLOv8 model. Validation on a dataset comprising four typical defect types (dark spots， holes， wrinkles， and scratches) demonstrated that the proposed model achieved improved feature extraction capability， parameter efficiency， and inference speed. Specifically， it attained a mean average precision (mAP₅₀) of 99.2%， with only 1.35 MB of parameter size and 3.07 GFLOPs of computational volume. On the RTX 4060 GPU， the inference frame rate reached 175 FPS， satisfying the requirements of a 600 m/min high-speed paper machine. |
| 查看全文
HTML
查看/发表评论 下载PDF阅读器 |