复杂环境下的植物病害识别新型研究-《计算机技术与发展》

文章信息/Info

Author(s):: LIU Zi-xi; LI Hong-xiang; FENG Ao; DU Zhi-xiong; NI Ming; School of Information Engineering,Sichuan Agricultural University,Yaan 625014,China

Keywords:: plant disease; RPN algorithm; Chan-Vese algorithm; transfer learning; ResNet-101

摘要:: 病害是威胁植物生长的主要因素之一,随着智慧农业的发展,实现在复杂环境下对植物病害的识别是更高效的防治植物病害的基础。针对复杂环境下的叶片识别问题,根据深度学习算法和迁移学习模型,构造出一种新型植物病害识别模型。首先使用复杂背景下的叶片数据集训练 RPN 算法( region proposal network,区域生成网络) 实现对叶片的检测定位,然后使用 Chan-Vese 算法分割图像得到包含病斑特征的叶片。最后,将分割后的叶片输入经过简单背景下病害叶片数据集训练后的迁移学习模型,实现复杂环境下的植物病害识别。在常见的植物叶片病害中,以褐斑病、霜霉病、灰霉病为例进行测试,通过测试结果表明该方法平均正确率为 90. 4% ,远高于传统的 ResNet-101 模型的正确率,在复杂环境下的植物病害识别应用上具有很好的实用性。

Abstract:: Disease is one of the main factors that threaten plant growth. With the development of smart agriculture, realizing the identification of plant diseases in complex environments is the basis for more efficient prevention and control of plant diseases. Aiming at the problem of leaf recognition in complex environments,we construct a new type of plant disease recognition model based on deep learning algorithm and transfer learning model. Firstly,the RPN algorithm is trained on the leaf data set in the complex background to realize the detection and positioning of the leaf, and then the Chan-Vese algorithm is used to segment the image to obtain the leaf containing the features of diseased spots. Finally,input the segmented leaves into the transfer learning model trained on the diseased leaf data set in a simple background to realize the recognition of plant diseases in a complex environment. Among the common plant leaf diseases,brown spot,downy mildew and gray mold are selected as examples for testing. It is showed that the average accuracy of the proposed method is 90. 4% ,which is much higher than that of the traditional ResNet-101 model. The accuracy rate is quite useful in the application of plant disease identification in complex environments.