深度学习训练数据分布对植物病害识别的影响研究

Impact of Training Data Distribution on Plant Diseases Recognition Based on Deep Learning

【目的】通过调节训练集内实验室场景图片与田间场景图片的分布,提高深度学习模型的准确度,以减少植物病害识别深度学习模型对田间场景数据的依赖。【方法】通过调节训练集内实验室场景图片和田间场景图片的分布,使用ResNeSt-50、VGG-16、ResNet-50等3种神经网络结构分别对训练得到的深度学习模型进行测试和比较,从而优化植物病害识别模型。【结果】在由一定数量的植物病害图像组成的训练集内,调节其中不同场景图片的分布会对模型的准确率产生影响。当训练集内的田间场景图片分布达30%时,模型准确率提升18%以上。在100%实验室场景图片的训练集内添加30%田间场景图片,可提升模型准确率17%以上;在100%田间场景图片的训练集内添加实验室场景图片,模型准确率随图片数量增加而提升,提升幅度为2%~4%。【结论】该方法适用于农业复杂环境下高准确度病害识别模型的快速建立,可减少深度学习模型对田间场景数据的依赖,缩短模型建立初期的田间数据采集周期,降低田间数据采集成本,促进人工智能技术在无人农场及智慧农业中更有效地运用。

【Objective】The study was carried out to improve the accuracy of the deep learning model through adjusting the distribution of training dataset of lab-condition and field-condition images,to reduce the dependence of plant diseases recognition models on field-condition data.【Method】The plant diseases recognition model was optimized through adjusting the distribution of images of lab-conditions and field-conditions in training datasets.Deep learning models of plant diseases trained by three artificial neural networks of ResNeSt-50,VGG-16 and ResNet-50 were tested and compared.【Result】In a training dataset composed of a certain number of plant disease images,it had an impact on the model accuracy through adjusting the distribution of images of different conditions.When the proportion of images of the fieldconditions reached 30%,the accuracy of the model was improved by more than 18%.Through adding field-conditions images at a number ratio of 30%into a training dataset composed of 100%lab-condition images,the accuracy of the model was improved by more than 17%.Through adding lab-conditions images into a training dataset composed of 100%field-condition images,the accuracy of the model was improved with the increasing number of images,and the improved ranges were between 2%and 4%.【Conclusion】This method is suitable for the rapid establishment of high-accuracy plant diseases recognition models in the complex agricultural environment.It could reduce the dependence of plant recognition models on field-condition images,shorten the field data collection cycle at the beginning of model establishment and reduce the cost of field-condition images collection.It promotes a more effective application of artificial intelligence in unmanned farms and smart agriculture.