| 摘要: |
| 为了及时准确的识别番茄叶片病害，提高番茄产量，提出了一种优化的Swin Transformer番茄病害识别方法，该模型利用Transformer的自注意力结构获得更加完备的番茄病害图像的高层视觉语义信息；结合Mixup混合增强算法，在预处理阶段对图像特征信息进行增强；并采用迁移学习在增强番茄叶片病害数据集上进行训练和优化Swin Transformer模型，以此实现精准的番茄叶片病害识别。结果表明:1)优化的Swin Transformer模型对番茄叶片病害识别准确率达到98.40%；2)在相同训练参数下，本研究模型比原Swin Transformer、VGG16、AlexNet、GoogLeNet、ResNet50、MobileNetV2、ViT和MobileViT模型准确率提高了0.70%~1.91%，且能快速收敛；3)本研究模型中加入的Mixup混合增强算法极大地提高了番茄叶片病害的识别准确率，比现有的常见方法性能更加优越，并且鲁棒性强。因此，本研究提出的新模型能够更加准确的识别番茄叶片病害。 |
| 关键词: Swin Transformer Mixup 数据增强 番茄病害识别 迁移学习 图像分类 |
| DOI：10.11841/j.issn.1007-4333.2023.04.07 |
| 投稿时间：2022-08-22 |
| 基金项目:国家自然科学基金项目(31870532)；湖南省自然科学基金项目(2021JJ31163)；湖南省教育科学“十三五”规划基金项目(XJK20BGD048) |
|
| Tomato leaf disease recognition based on an optimized Swin Transformer |
|
LIU Yongmin,LIU Hanlin,SHI Tingting,OUYANG Jinyi,HUANG Hao,XIE Tieqiang
|
| (School of Computer and Information/ Research Center of Smart Forestry Cloud, Central South University of Forestry and Technology, Changsha 410004, China) |
| Abstract: |
| To timely and accurately recognize tomato leaf diseases, an optimized Swin Transformer method is proposed. A self-attention structure of Transformer is used to obtain the more complete high-level visual semantic information of tomato disease images. Mixup hybrid enhancement algorithm is combined to enhance image feature information in the preprocessing stage; In addition, transfer learning is adopted to train and optimize Swin Transformer model on enhanced tomato leaf disease data set, so as to achieve accurate tomato leaf disease identification. The results showed that: 1)With the optimized Swin Transformer model, the identification accuracy of tomato leaf disease reached 98. 40%; 2)Under the same training parameters, the accuracy of this model is 0. 70% to 1. 91% higher than that of the original Swin Transformer, VGG16, AlexNet, GoogLeNet, ResNet50, MobileNetV2, ViT and MobileViT models, and it could also rapidly convergence; 3)The Mixup hybrid enhancement algorithm added to the model greatly improved the identification accuracy of tomato leaf diseases, and had better performance and strong robustness than the existing common classical methods. Therefore, the new model proposed in this study can identify tomato leaf diseases more accurately. |
| Key words: Swin Transformer Mixup data augmentation tomato disease recognition transfer learning image classification |
