| 王水明,李鹏飞,林杰,李浥尘,唐天明,陈前进,汪涵,魏宁丰.基于轻量化目标检测模型的再造烟叶表面缺陷识别方法[J].中国造纸,2026,45(5):199-207 |
二维码（扫一下试试看！） |
| 基于轻量化目标检测模型的再造烟叶表面缺陷识别方法 |
| Surface Defects Identification Method for Reconstituted Tobacco Based on Lightweight Object Detection Model |
| 收稿日期：2026-01-20 修订日期：2026-02-06 |
| DOI：10.11980/j.issn.0254-508X.2026.05.023 |
| 关键词: 再造烟叶 缺陷检测 深度学习 双注意力机制 模型轻量化 |
| Key Words:reconstituted tobacco defect detection deep learning dual-attention mechanism lightweight model |
| 基金项目:湖北省自然科学基金面上项目（2023AFB878）；湖北省自然科学基金青年项目（2024AFB259）。 |
| 作者 | 单位 | 邮编 | | 王水明* | 1湖北中烟工业有限责任公司,湖北武汉,430040
2湖北新业烟草薄片开发有限公司,湖北武汉,430056 | 430056 | | 李鹏飞* | 1湖北中烟工业有限责任公司,湖北武汉,430040
2湖北新业烟草薄片开发有限公司,湖北武汉,430056 | 430056 | | 林杰 | 1湖北中烟工业有限责任公司,湖北武汉,430040
2湖北新业烟草薄片开发有限公司,湖北武汉,430056 | 430056 | | 李浥尘 | 1湖北中烟工业有限责任公司,湖北武汉,430040
2湖北新业烟草薄片开发有限公司,湖北武汉,430056 | 430056 | | 唐天明 | 1湖北中烟工业有限责任公司,湖北武汉,430040
2湖北新业烟草薄片开发有限公司,湖北武汉,430056 | 430056 | | 陈前进 | 1湖北中烟工业有限责任公司,湖北武汉,430040
2湖北新业烟草薄片开发有限公司,湖北武汉,430056 | 430056 | | 汪涵 | 3武汉纺织大学机械工程与自动化学院,湖北武汉,430200 | 430200 | | 魏宁丰 | 3武汉纺织大学机械工程与自动化学院,湖北武汉,430200 | 430200 |
|
| 摘要点击次数: 600 |
| 全文下载次数: 110 |
| 摘要:针对造纸法再造烟叶表面缺陷人工检测存在的低效率、高误检、难定位等问题，本研究提出一种基于轻量化目标检测模型的缺陷自动识别方法。首先，搭建基于双线阵相机与辅助光源的缺陷检测平台采集缺陷图像；然后，采用图像批量预处理与数据增强方法，对数据集进行扩展与优化，并采用人工与自动标注相结合的方式，依据缺陷特点进行多类别标注，获得足量高精度模型训练数据；最后，构建基于双注意力机制和双检测头的轻量化目标检测模型YOLO-D。结果表明，该模型的精确率为87.7%，计算量仅为20.8 GFLOPs，相比YOLOv8模型精确率提升了2.2%，计算量降低了26.8%，实现了再造烟叶表面缺陷的快速、准确识别。 |
| Abstract:To address the issues of low efficiency， high false detection， and difficult localization in manual inspection of surface defects in paper-making reconstituted tobacco， this paper proposed an automatic defect recognition method based on a lightweight object detection model. First， a defect detection platform based on dual-line-scan cameras and auxiliary lighting was constructed to capture defect images. Subsequently， batch image preprocessing and data augmentation methods were proposed to expand and optimize the dataset. A combined manual and automatic annotation approach was adopted to perform multi-category labeling according to defect characteristics， obtaining sufficient high-precision training data. Finally， a lightweight object detection model named YOLO-D was developed， incorporating a dual-attention mechanism and a dual-detection-head structure. The results showed that the model achieved a precision of 87.7% with a computational load of only 20.8 GFLOPs. Compared to the YOLOv8 model， precision was improved by 2.2% and computational load was reduced by 26.8%. The proposed method enabled fast and accurate identification of surface defects in reconstituted tobaccos. |
| 查看全文
HTML
查看/发表评论 下载PDF阅读器 |
|
|
|