基于深度学习的农作物病害检测

Crop disease detection based on deep learning

针对在大规模农业种植中传统人工农作物病虫害预防和治理上常存在的问题,应用深度学习算法来进行农作物病害的检测.对47 637张图片进行病害识别检测,数据包含10个物种(主要农作物有番茄、土豆、玉米等),27种病害,总共61个分类标签.采用目前流行的深度网络结构如Vgg-16,ResNetV1-101和InceptionV4等6种模型对图像进行特征抽取.采用交叉熵和正则化项组成损失函数进行反向传播调整,对数据集进行4种不同情况的划分;并且使用了初始化和迁移训练两种训练方式,分别对6种网络架构在不同学习率下进行试验比较.结果表明:采用初始化训练对61类病害情况的最高识别准确率为84.6%;而在迁移训练中,使用合适的学习率训练,最高识别准确率达到86.1%;对3类疾病程度分类准确率为87.4%,对28种病害类型分类准确率为98.2%,对10类物种识别分类准确率为99.3%.

To solve the problems in traditional artificial crop disease prevention and reduction in large scale agriculture production,the deep learning algorithm was used to detect and recognize crop diseases. The 47 637 images were detected for disease identification with 10 species of tomatoes,potatoes and corn,27 diseases and total 61 classes. Six currently popular deep network structures of Vgg-16,ResNetV1-101,InceptionV4,etc. were applied to perform the feature extraction. The loss function with cross entropy and regularization was adopted to conduct back propagation. The data set was divided according to four different cases,and the initialization and the transfer learning were used in the training procedure. Six network structures with different learning rates were compared in the experiment. The results show that the highest recognition rate of all 61 classes is 84.6% by the initialization learning while the highest rate is 86.1% by the transfer learning and appropriate learning rate. For three statuses of diseases,the recognition rate is 87.4%. For 28 diseases,the recognition rate is 98.2%,and the recognition rate of the 10 types of diseases is 99.3%.