基于迁移学习的水稻病虫害识别

Identification of rice diseases and insect pests using transfer learning

关键词:  水稻病虫害识别  深度学习  迁移学习  PyTorch  ResNet

Key words:  identification of rice diseases and insect pests  deep learning  tranainiisfer learning  PyTorch  ResNet

摘要:

【目的】水稻病虫害是引起水稻减产的重要因素。准确地识别水稻病虫害类型，及 时采取有效的针对性预防措施，有助于避免因水稻减产带来的经济损失。然而，聚焦于人 脸和花草等常见事物的识别技术，在农业领域特别是水稻病虫害识别领域应用较少，而 目前已有的水稻病虫害识别研究存在数据量小和数据种类不够丰富等问题。【方法】文 章搜集了2.0372 万张水稻病虫害图片，并以此构建了完整的水稻病虫害识别数据集，基 于迁移学习的思想，在ResNet50 的预训练模型基础上构建了一个针对16 种主要水稻病 虫害识别的深度模型。同时，考虑实际应用的需要，搜集了9 928 张其他图片（包括人 像、汽车等），结合9 675 张水稻病虫害图片，构建了一个二分类数据过滤模型，以此来 避免非水稻病虫害图片被识别为某一类病虫害的不合理结果。【结果】有预训练模型验 证结果的top-1 准确率达到了95.23%，F1 系数为77.83%，相较无预训练模型top-1 准确 率提升了24.51%，F1 系数提升了56.66%。数据过滤模型的过滤准确度达到了99.60%。 【结论】基于迁移学习的水稻病虫害识别模型，使水稻病虫害识别结果更加准确。非水稻病 虫害过滤模型，有效地解决了实际应用中非水稻病虫害图片被错分为某一类水稻病虫害的 问题。

Abstract:

［Purpose］Rice diseases and insect pests are important factors causing rice yieldlosses. Accurate recognition of rice diseases and insect pests facilitates timely preventive measuresto avoid economic losses. However，the existing identification technologies mainly focus oncommon objects. It is seldom used in agriculture，especially in rice diseases and insect pestsidentification. The existing researches on rice diseases and insect pest identification have problemsof small data volume and insufficient data types.［Method］In this paper，20 372 pictures ofrice insect pests and diseases were collected to build a complete dataset for rice diseases andinsect pests identification. Based on transfer learning，a deep recognition model for 16 major rice diseases and insect pests was built on ResNet50 pre-training model. In addition，for practicalapplication，with 9 928 other pictures（including portraits，cars，etc.）and 9 675 pictures ofrice diseases and insect pests，this paper constructed a binary filtering model based on ResNet50pre-training model to avoid the pictures of others from being misclassified into a certain typeof rice diseases and insect pests.［Result］The top-1 accuracy of the rice diseases and insectpests classification model achieves 95.23%，the top-5 accuracy achieves 96.33%，the kappacoefficient is 0.936 8 and the F1 coefficient is 77.83%. Compared with the non-pre-trainingmodel，the accuracy of top-1 increased by 24.51%，the top-5 accuracy increased by 22.02%，the kappa coefficient increased 0.349 2 and the F1 coefficient increased by 56.66%. It shows thattransfer learning does contribute to the outstanding performance of the rice diseases and insectpests identification. And the filtering model’s top-1 accuracy is 99.60%.［Conclusion］20 372pictures of rice diseases and insect pests were collected to build a complete dataset through fieldand network，including 16 common rice diseases and insect pests. The rice diseases and insectpests identification model based on transfer learning makes the identification of rice diseasesand insect pests more accurate. In addition，the filtering model effectively solves the problem ofnon-rice pictures being misclassified into a certain type of rice diseases and insect pests.