Received  25 July, 2017    Accepted  19 February, 2018
 

Maize (Zea mays L.), also known as corn, is the third most cultivated crop in the world.  Northern corn leaf blight (NCLB) is a globally devastating maize foliar disease caused by Setosphaeria turcica (Luttrell) Leonard and Suggs.  Early intelligent diagnosis and warning is an effective and economical strategy to control this disease.  Today, deep learning is beginning to play an essential role in agriculture.  Notably, deep convolutional neural networks (DCNN) are amongst the most successful machine learning techniques in plant disease detection and diagnosis.  Our study aims to identify NCLB in the maize-producing area in Jilin Province based on several DCNN models.  We established a database of 985 leaf images of healthy and infected maize and applied data augmentation techniques including image segmentation, image resizing, image cropping, and image transformation, to expand to 30 655 images.  Several proven convolutional neural networks, such as AlexNet, GoogleNet, VGG16, and VGG19, were then used to identify diseases.  Based on the best performance of the DCNN pre-trained model GoogleNet, some of the recent loss functions developed for deep facial recognition tasks such as ArcFace, CosFace, and A-Softmax were applied to detect NCLB.  We found that a pre-trained GoogleNet architecture with the Softmax loss function can achieve an excellent accuracy of 99.94% on NCLB diagnosis.  The analysis was implemented in Python with two deep learning frameworks, Pytorch and Keras.  The techniques, training, validation, and test results are presented in this paper.  Overall, our study explores intelligent identification technology for NCLB and effectively diagnoses NCLB from images of maize.

准确、及时地对草莓种植过程中的病害进行分类，可以帮助种植者对其进行及时的处理，从而减少损失。但真实种植环境下的草莓病害分类面临着严峻的挑战，包括复杂的种植环境，多种差异较小的病害类别等。尽管最近基于深度学习的移动视觉技术在克服上述问题方面取得了一些成功，但对多地域、多空间、多时间的草莓病害分类需求而言，一个关键问题是如何构建一种无损、快速、便捷的方法提高草莓病害识别的效率。我们开发并评估一种快速，低成本的系统，用于对草莓种植中的病害进行分类。这涉及设计一个易于使用的基于云服务的草莓病害识别系统，并结合我们提出的新颖的自监督多网络协作的分类模型，它由定位网络、反馈网络和分类网络组成，以识别草莓常见病害的类别。该模型借助新颖的自我监督机制，可以有效地识别草莓病害图像中的病害区域，而不需要边界框等标注。使用准确率，精确率，召回率和值来评估分类效果，测试集的结果分别为92.48％，90.68％，86.32%和88.45%。与流行的卷积神经网络和其他五种方法相比，该网络实现了更好的病害分类效果。目前，系统的客户端（小程序）已上线微信平台。小程序在实际测试中分类效果良好，验证了系统的可行性和有效性，能够为草莓病害识别的智能化研究与应用提供参考。