| 摘 要: 诊断与识别植物叶片的病虫害是农业生产中的一大难题。为了解决西瓜叶片病虫害的诊断与识别问题,方便瓜农评估西瓜叶片的健康状况,提出了一种先分割、后识别的西瓜叶片病害识别算法。该算法首先采用UNet模型对叶片进行分割,其次使用Swin-Transformer模型进行病虫害识别。通过在自建的西瓜叶片数据集上进行对比实验,文章所提算法识别准确率达到92.9%,相比直接在原始图像上使用Swin-Transformer模型进行病虫害识别,准确率提高了3.2%。实验结果表明,使用分割后的图像进行病虫害分类可以显著提高识别准确率。 |
| 关键词: UNet;Swin-Transformer;语义分割;病虫害识别 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 2021年湖南农业大学大学生创新创业训练项目(XCX2021003) |
|
| Research on Watermelon Leaf Disease Identification Based on UNet+Swin-Transformer |
|
XIANG Yujie, XIANG Yuanping
|
(College of Inf ormation and Intelligence, Hunan Agricultural University, Changsha 410128, China)
2014176995@qq.com; 13619652@qq.com
|
| Abstract: Diagnosis and identification of plant leaf diseases and pest are major challenges in agricultural production. In order to solve the problems of diagnosis and identification of watermelon leaf disease and pest, and to facilitate the evaluation of the health status of watermelon leaves by farmers, this paper proposes a watermelon leaf disease identification algorithm with segmentation first and recognition followed. Firstly, UNet model is used for leaf segmentation, and then Swin Transformer model is used for pests and diseases identification. Through comparative experiments on a self-built watermelon leaf dataset, the proposed algorithm achieves a recognition accuracy of 92.9% , which is 3.2% higher than the one that uses Swin Transformer model directly on the original image for pest and disease identification. The experimental results show that using segmented images for pest and disease classification can significantly improve recognition accuracy. |
| Keywords: UNet; Swin-Transformer; semantic segmentation; identification of pests and diseases |
