基于迁移学习的棉花叶部病虫害图像识别

针对传统图像识别方法准确率低、手工提取特征等问题,该研究以棉花叶部病虫害图像为研究对象,利用迁移学习算法并辅以数据增强技术,实现棉花叶部病虫害图像准确分类。首先改进AlexNet模型,利用PlantVillage大数据集训练取得预训练模型,在预训练模型上使用棉花病虫害数据微调参数,得到平均测试准确率为93.50%;然后使用数据增强技术扩充原始数据集,在预训练模型上再训练,得到最终平均测试准确率为97.16%。相同试验条件下,该研究方法较支持向量机(Support Vector Machine,SVM)和BP(Back Propagation,BP)神经网络以及深度卷积模型(VGG-19和GoogLeNet Inception v2)分类效果更好。试验结果表明,通过迁移学习能把从源领域(PlantVillage数据集)学习到的知识迁移到目标领域(棉花病虫害数据集),数据增强技术能有效缓解过拟合。该研究为农作物病虫害识别技术的发展提供了参考。

Application of Computer Vision Technology in Agriculture

With the development of image processing technology and computer, computer vision technology has been widely used in the production of agriculture,and has made many important achievements. This paper reviews its-research progress on diagnosis of agricultural products, water diagnosis, weed identification,product quality testing and grading, agricultural picking and sorting and other as- pects, and finally put forward its existing problems and prospects for the future.