基于迁移学习的葡萄叶片病害识别及移动端应用

为解决已有的卷积神经网络在小样本葡萄病害叶片识别的问题中出现的收敛速度慢,易产生过拟合现象等问题,提出了一种葡萄叶片病害识别模型(Grape-VGG-16,GV),并针对该模型提出基于迁移学习的模型训练方式。将VGG-16网络在ImageNet图像数据集上学习的知识迁移到本模型中,并设计全新的全连接层。对收集到的葡萄叶片图像使用数据增强技术扩充数据集。基于扩充前后的数据集,对全新学习、训练全连接层的迁移学习、训练最后一个卷积层和全连接层的迁移学习3种学习方式进行了试验。试验结果表明,1)迁移学习的2种训练方式相比于全新学习准确率增加了10～13个百分点,并在仅训练25轮达到收敛,该方法有效提升了模型分类性能,缩短模型的收敛时间;2)数据扩充有助于增加数据的多样性,并随着训练次数的增加,训练与测试准确率同步上升,有效缓解了过拟合现象。在迁移学习结合数据扩充的方式下,所构建的葡萄叶片病害识别模型(GV)对葡萄叶片病害的识别准确率能达到96.48%,对健康叶、褐斑病、轮斑病和黑腐病的识别准确率分别达到98.04%、98.04%、95.83%和94.00%。最后,将最终的研究模型部署到移动端,实现了田间葡萄叶片病害的智能检测,为葡萄病害的智能诊断提供参考。

Recognition of grape leaf diseases and mobile application based on transfer learning

Abstract: Pests and diseases have posed a severe threat to the production of grapes, causing irreversible damage on the growth cycle of grapes. The treatments vary significantly in the types of grape diseases. Therefore, it is highly demanding to detect the diseases of grapes timely, while accurately identify the types of diseases to avoid the spread of diseases. In the early days, the diseases of grapes were mainly identified by experienced grape and pest experts recognizing the grape leaves manually, which is time-consuming and laborious. It is necessary to develop a grape disease recognition using computer science technology. Convolutional neural network (CNN) has been used to automatically recognize agricultural pests in the field of artificial intelligence (AI) in recent years. However, the general CNN cannot be well trained to achieve high accuracy, due mainly to few large datasets of diseased grape leaves. In this study, a new recognition network of grape leaf disease was proposed using CNN. Transfer learning and data enhancement were also applied for the new network to enlarge the small dataset of grape leaves. The new network was named Grape-VGG-16 (GV for short) using VGG-16 network (one of the most classical CNN). GV generally contained 5 convolutions, 2 full connection 1 flatten, 1 dropout, and 1 SoftMax layer. The knowledge was first learned from the ImageNet image dataset using transfer learning, and then be transferred to the GV in the disease recognition of grape leaf. The output number was modified to 4 for the last fully connected layer using fine tuning, representing the 4 types of grapes leaves (healthy, brown spot, wheel spot, and black rot). A TensorFlow framework was selected to verify the network in the experiments using Python programming language. 300 images were collected for each leaf category (the total number of images was 1200), 250 of which were randomly selected as the training set, and 50 of which as the testing set. Then the dataset of images was augmented under the operations, such as random rotation, random horizontal and vertical translation, as well as horizontal flipping on the image. Finally, the augmented dataset was used to train the GV network, which was uniformly scaled to 224×224 pixels. 18 groups of experiments were performed on a computing server with Nvidia GPU using a combination of three learning (new learning, transfer learning with training the full connection layer, as well as transfer learning with training both the last convolution and full connection layer), two dataset with/without augmentation, and three learning rates (0.01, 0.001, and 0.0001). The experimental results demonstrated that: 1) Transfer learning improved the model accuracy, while shortened the convergence time of the model; 2) Data augmentation greatly increased the diversity of data, while effectively alleviated the over-fitting. Consequently, the GV network achieved an overall recognition accuracy of 96.48% under both transfer learning and data augmentation. The recognizing accuracies of healthy leaves, brown spot, wheel spot and black rot were 98.04%, 98.04%, 95.83%, and 94.00%, respectively. The GV network was also embedded in the mobile application, and then deployed on the mobile terminal, for the detection and recognition of grape leaf diseases in the field. The finding can offer a potential promising reference for the intelligent diagnosis of grape leaf diseases in fruit production.