数据集对基于深度学习的作物病害识别有效性影响

基于深度学习的作物病害自动识别已成为农业信息化领域新的研究热点,为探究数据集的大小和质量对基于深度学习的作物病害识别有效性的影响,研究不同数据集训练得到的模型识别效果并进行了分析。以338张玉米病害数据集为例,对其进行数据增强、移除背景和细分割划分等处理,设计5个AlexNet框架的CNN网络模型并利用不同类型的数据集进行测试,使用十倍交叉法验证识别结果。试验结果表明:使用不同数量和等级的数据训练后的模型识别准确率分别为56.80%、78.30%、80.50%、89.30%和81.00%。在获得每个网络的最终精度后,挑选出识别错误的图像进行分析,结合前人的研究结果,得出影响深度神经网络用于作物病害识别有效性的9个主要因素。数据集对基于深度学习的作物病害识别有效性的影响因素主要分为带注释的数据集大小、叶片症状代表性、协变量转移、图像背景情况、图像数据获取过程、症状分割、特征多变性、并发性病害以及症状的相似性等这9类,该试验能够为深度学习技术田间病害识别的实际应用中提供依据和指导。

Influence of data sets on the effectiveness of crop disease recognition based on deep learning

Automatic recognition of crop disease based on deep learning has become a new research area in agricultural informatization.In order to explore the influence of the size and quality of data sets on the effectiveness of crop disease recognition based on deep learning,the model recognition effects obtained by different training data sets were studied and analyzed.Corn disease data set with the number of 338 was taken as sample,which was augmentation enhanced,background-removed and divided into smaller images containing individual lesions or localized symptom regions.Five CNN network models of AlexNet framework were designed and they were trained with different types of data sets and tested respectively.The final accuracy for each network was obtained using a 10-fold cross-validation approach.The experiments illustrated that the recognition models get different recognition accuracy of 56.80%,78.30%、80.50%、89.30%and 81.00%.The characteristics of the images of misrecognition by the CNNs were picked out and analyzed carefully,previous studies were analyzed as well,and nine factors of data set were identified as having the most impact on the crop disease recognition results.Analyzed comprehensively,the annotated data set size,leaf symptoms representative,covariate transfer,image background,image data acquisition condition,segmentation,feature variability,concurrent disease symptoms and symptoms of similarity are the influencing factors on the effectiveness of deep learning in crop disease recognition.This experiment can provide the basis and guidance for the application of deep learning technology in field disease recognition.