面向不平衡数据的木薯叶部病害图像识别方法  

Cassava leaf disease image recognition method for imbalanced data

作　　者：王丹阳 梁伟红[1] 李玉萍[1] 黄贵修 Wang Danyang;Liang Weihong;Li Yuping;Huang Guixiu(Institute of Scientific and Technical Information,Chinese Academy of Tropical Agricultural Sciences/Key Laboratory of Practical Research on Tropical Crops Information Technology in Hainan,Haikou,571101,China;Environment and Plant Protection Institute,Chinese Academy of Tropical Agricultural Sciences,Haikou,571101,China)

机构地区：[1]中国热带农业科学院科技信息研究所/海南省热带作物信息技术应用研究重点实验室,海口市571101 [2]中国热带农业科学院环境与植物保护研究所,海口市571101

出　　处：《中国农机化学报》2025年第3期101-107,123,共8页Journal of Chinese Agricultural Mechanization

基　　金：中央级公益性科研院所基本科研业务费专项(1630072023005);海南省自然科学基金青年基金项目(323QN300)。

摘　　要：为提高产地环境下木薯叶部病害自动识别的准确性,解决病害图像低对比度和数据长尾分布问题,建立一种深度学习模型SwinTFCC用于木薯叶部病害识别。该模型采用Swin Transformer作为骨干网络,借助Swin Transformer的自注意力机制和层级结构关注局部与全局特征,使其对复杂背景病害识别具有鲁棒性;将最后一层特征输入特征簇压缩模块,以映射稀疏特征簇为稠密特征簇,减少长尾分布中样本少的类别稀疏特征簇跨越决策边界导致分类错误情况;并采用迁移学习在木薯叶部病害图像数据集上进行训练,以提升木薯叶部病害识别性能。试验结果表明,模型的F 1值达到90.74%,较其他主流模型提升8.04%~19.3%。所采用的方法在小规模不平衡数据集上取得较好效果,验证模型的有效性,为木薯叶部病害自动精准识别提供技术支撑。To improve the accuracy of automatic cassava leaf disease recognition in production environments and address issues such as low-contrast disease images and long-tail data distribution,this paper proposes a deep learning model,SwinTFCC,for the recognition of cassava leaf diseases.This model employs the Swin Transformer as the backbone network,utilizing its self-attention mechanism and hierarchical structure to focus on local and global features for robust disease recognition in complex backgrounds.The features from the last layer are then input into a feature cluster compression module to map sparse feature clusters into dense ones,reducing classification errors caused by sparse feature clusters of underrepresented classes crossing decision boundaries in the long-tailed distribution.The model is trained on a cassava leaf disease image dataset using transfer learning to enhance recognition performance.The experimental results indicate that the proposed model achieves an F1 score of 90.74%,improving by 8.04% to 19.3% compared with other mainstream models.In this study,the method performs well on a small-scale imbalanced dataset,confirming the model's effectiveness and providing technical support for the automatic and precise recognition of cassava leaf diseases.

关 键 词：木薯叶部 病害识别 图像识别 迁移学习 不平衡数据 

分 类 号：S435.33[农业科学—农业昆虫与害虫防治] TP391.4[农业科学—植物保护]