摘要
病虫害影响水稻质量和产量,快速、准确地检测出水稻病虫害有利于及时防治。针对传统图像识别方法存在特征提取繁琐、识别率低以及对田间环境下的作物病虫害识别困难等问题,本文提出一种以DenseNet121为基础网络,结合迁移学习与坐标注意力机制的水稻病虫害识别模型。该模型引入坐标注意力学习图像特征的通道间关系和空间位置的重要性以增强模型的特征提取能力,采用迁移学习策略训练模型以缓解模型在小数据集上的过拟合现象、减小计算资源以及提升模型的识别性能。利用从田间复杂环境收集的水稻病虫害数据集,对该模型与ResNet50、Xception、InceptionV3、InceptionResNetV2及原DenseNet121等卷积神经网络模型的识别效果进行比较,结果表明,该模型能有效识别出水稻常见8种病虫害和健康植株,识别准确率达到98.95%,模型参数量仅为7.23 M,识别效果优于其他模型。这可为田间环境下的其他作物病虫害识别提供参考。
Rice diseases and pests have important impacts on quality and yield of rice.Therefore,it is vital to identify the types of rice diseases and pests quickly and accurately for controlling them timely and accurately.In view of the tedious extraction of image features,low accuracy and difficult recognition in field environment of traditional crop disease and pest identifying methods,we established a model to identify rice diseases and pests based on the convolutional neural network DenseNet121 and developed by transfer learning and coordinate attention(CA)mechanism.This model embedded the CA module to learn the importance of inter-channel relationship and spatial location for input features to enhance the ability of feature extraction,and used fransfer learning strategy in model training to alleviate the phenomenon of overfitting on small data set,save more computational power and improve the identification accuracy.The open dataset of rice diseases and pests collected from the field with complex environment conditions was used to compare identification effect of the proposed model with the ResNet50,Xception,InceptionV3,InceptionResNetV2,and original DenseNet121 models.The results showed that the proposed model could effectively identify eight common rice diseases and pests and healthy rice plants with the recognition accuracy of 98.95%but the model parameter as only 7.23 M,which was obvious better than other models.This research could provide references for recognition of other crop diseases and pests in field environment.
作者
陈浪浪
张艳
Chen Langlang;Zhang Yan(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;Center of Nondestructive Testing for Agricultural Products,Guiyang University,Guiyang 550005,China)
出处
《山东农业科学》
北大核心
2023年第5期164-172,共9页
Shandong Agricultural Sciences
基金
国家自然科学基金项目(62141501)。
作者简介
陈浪浪(1996-),男,硕士研究生,研究方向为图像处理、作物病害的无损检测。E-mail:wychenlanglang@163.com;通信作者:张艳(1977-),女,博士,教授,硕士生导师,主要从事生物信息无损检测、激光雷达方面的研究。E-mail:Eileen_zy001@sohu.com。
