| 刘鑫,周强,陈雅鑫,张晓辉.基于重参数化与多查询注意力的轻量化纸病检测模型[J].中国造纸学报,2025,40(4):186-194 |
二维码（扫一下试试看！） |
| 基于重参数化与多查询注意力的轻量化纸病检测模型 |
| Lightweight Paper Defect Detection Model Based on Re-parameterization and Mutile Query Self Attention |
| 投稿时间：2024-12-17 修订日期：2024-12-23 |
| DOI：10.11981/j.issn.1000-6842.2025.04.186 |
| 中文关键词: 纸病检测 RepLite-YOLO 重参数化 多查询注意力 NMS-Free |
| Key Words:paper defect detection RepLite-YOLO re-parameterization mutile query self attention NMS-Free |
| 基金项目:陕西省科技厅工业领域一般项目（2024GX-YBXM-544）。 |
|
| 摘要点击次数: 33 |
| 全文下载次数: 22 |
| 中文摘要: |
| 为了解决现有深度学习算法在纸病检测中难以兼顾检测精度和速度，导致基于机器视觉的纸病检测系统实用性较差的问题，本研究提出了一种基于重参数化与多查询注意力的轻量化纸病检测模型RepLite-YOLO。首先，设计了轻量级的部分深度可分离卷积，降低了模型的复杂度；在此基础上，构建了RCNXC2f模块，以增强模型对多尺度特征的提取能力，并使用重参数化减少了模型推理时的计算量；之后，在模型主干特征提取网络尾部增加C2PMQA模块，以提升模型对全局特征的提炼能力；最后，设计了轻量化检测头模块，并引入无非极大值抑制训练策略，有效提升了纸病检测的实时性。在自建纸病数据集上的研究结果表明，本研究提出的RepLite-YOLO模型的平均正确率为99.2%，参数量仅为1.35 MB，在NIVIDA 4060平台上，检测分辨率为512×512的纸病图像帧率达到了175帧。RepLite-YOLO模型不仅能够保证纸病检测的精度，而且在实时性方面优于其他纸病检测模型。 |
| Abstract: |
| To address the challenges in existing deep learning algorithms， which struggle to balance detection accuracy and speed in paper defect detection， leading to the limited practicality of machine vision-based systems， this paper proposed a lightweight paper defect detection model named RepLite-YOLO based on re-parameterization and mutile query self attention. First， a lightweight partially depthwise separable convolution was designed to reduce model complexity. On this basis， the RCNXC2f module was constructed to enhance the model’s capability of extracting multi-scale features， with re-parameterization employed to reduce computational costs during inference. Subsequently， the C2PMQA module was integrated into the tail of the backbone network to improve the refinement of global features. Finally， a lightweight detection head module was developed， incorporating a training strategy without Non-Maximum Suppression (NMS) to effectively enhance real-time detection performance. Experimental results on a self-constructed paper defect dataset demonstrated that the proposed RepLite-YOLO achieves a mean Average Precision (mAP) of 99.2%， with a model size of only 1.35 MB. On the NVIDIA 4060 platform， the model achieved a frame rate of 175 FPS for 512×512 resolution paper defect images. RepLite-YOLO model not only ensured the precision of paper defect detection， but also outperformed other paper defect detection models in terms of real-time performance. |
| 查看全文 查看/发表评论 下载PDF阅读器 HTML |
