|
| 基于ResNet-AutoMix的仓储烟草品质分级 |
| ResNet-AutoMix based Quality Grading Algorithm for Warehousing Tobacco |
| 投稿时间：2023-12-01 修订日期：2024-03-01 |
| DOI： |
| 中文关键词: 烟草品质检测 不平衡数据 数据增强算法 |
| 英文关键词: Tobacco quality assessment Long-tail data Data enhancement algorithm |
| 基金项目: |
|
| 摘要点击次数: 70 |
| 全文下载次数: 0 |
| 中文摘要: |
| 烟草智慧养护系统可通过视觉技术实时监测烟草的品质等级，从而大规模精准调配仓储原料。当前绝大多数基于视觉的烟叶品质分级方法仅针对单片烟叶，难以满足系统判定烟草整体质量的需求。本研究针对烟草仓储过程中各级仓储原料数据量不平衡且类间相似度过高的问题，首次提出一种针对大规模仓储烟草集群的品质分级算法，将ResNet-34网络应用到大规模烟草分级任务。由于传统网络存在收敛性差、无法准确识别小样本数据的问题，因此我们在训练期间采用AutoMix数据增强算法克服该问题。实验结果表明，数据增强方法使准确率提升近4%，所提方法在测试集上的平均分级准确率达到了92.42%，实现了参数量和准确性的最优解。 |
| 英文摘要: |
| The Intelligent Maintenance System of Tobacco with computer vision technology can monitor the quality of tobacco in real time, enabling large-scale precision allocation of storage materials. Most computer- vision based tobacco leaf quality grading methods only focus on single tobacco leaf, which can’t meet the demand of the system for determining the overall leaves quality. To tackle the issue that data imbalance and high inter-class similarity in storage tobacco, our work firstly propose a quality grading algorithm for large-scale storage tobacco groups, adopting the ResNet-34 network to large-scale tobacco grading tasks. Due to the poor convergence and inability to accurately identify few-shot data exhibited by traditional networks, AutoMix is adopted in the training process to overcome this issue. Experimental results demonstrate that the data augmentation method improves accuracy by nearly 4%. The proposed method achieves an average grading accuracy of 92.42% on the test set, thereby achieving the optimal balance between parameter quantity and accuracy. |
|
View Fulltext
查看/发表评论 下载PDF阅读器 |
| 关闭 |
