| 摘 要: 针对自然条件下马铃薯块茎品种识别效率低、准确性差及化学分析法的不足,提出基于改进 ResNet50模型的识别方法。该方法通过在主干网络引入注意力模块,调整网络结构,采用 AdamW 优化器,加入迁移学习等改进措施,有效提升了模型性能。实验基于69个品种、30930张图片的数据集,最终识别准确率达99.42%,精确率、召回率、F1值也表现优异,相比 MobileNet_V2、GoogLeNet和ResNet50有显著提高,为马铃薯产业智能化管理提供了可靠技术支撑。 |
| 关键词: 深度学习 ResNet50 马铃薯品种识别 注意力机制 迁移学习 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 国家自然科学基金项目(32360437);甘肃省高等学校产业支撑计划项目(2021CYZC-57) |
|
| Research on Potato Recognition and Classification Method Based on the Improved ResNet50 |
|
WANG Jianwen1,LIU Chengzhong1, HAN Junying1, QU Yaying2, MA Baixiong1
|
(1.College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China;
2.Potato Research Institute, Gansu Academy of Agricultural Sciences, Lanzhou 730070, China)
1341673306@qq.com; liucz@gsau.edu.cn; hanjy@gsau.edu.cn; 605314800@qq.com; 3126540493@qq.com
|
| Abstract: To address the low efficiency and accuracy of potato tuber variety recognition under natural conditions, as well as the limitations of chemical analysis methods, this study proposes an identification approach based on an improved ResNet50 model. The method incorporates several enhancements: introducing an attention module into the backbone network, adjusting the network structure, adopting the AdamW optimizer, and integrating transfer learning.These improvements significantly boost model performance. Experiments were conducted on a dataset comprising 30 930 images across 69 potato varieties. The final recognition accuracy reached 99.42% , with outstanding results in precision, recall, and F1-score. Compared to MobileNet_V2, GoogLeNet, and the original ResNet50, the proposed model demonstrates substantial improvements, offering reliable technical support for intelligent management in the potato industry. |
| Keywords: deep learning ResNet50 potato variety identification attention mechanism transfer learning |
