Research On Deep Learning Based Recognition Of Tea Diseases For Android

China is a large exporter and consumer of tea,and 19 provinces(autonomous region)produce tea in the country,the planting and processing of tea directly influence local economic development.Large-scale cultivation of tea trees is accompanied by the breeding of large-scale infectious diseases of tea.Tea infections directly affect the yields and quality of tea.Therefore,it is particularly important to recognize and prevent the diseases of tea in time.Due to the wide variety of diseases of tea and different symptoms in different stages,the recognition mainly depends on the experience of tea farmers at present,which is subjective to some extent.With the development of AI technology,the agricultural related field has walked into intelligent modes.In the aspect of recognition of crops diseases,image recognition technology based on deep learning has gradually replaced the traditional way relying on manual design of feature extraction algorithm.In recent years,deep learning has developed rapidly.With the continuous optimization of the structure of deep learning models,the recognition of crops diseases can even be carried out locally in mobile devices,which provides great convenience to farmers.In view of the above situation,this thesis carrys out a research on the recognition of tea diseases for Android,which is based on the feature learning method of deep convolutional neural network.The main works and results are shown as follows:(1)In view of the scarcity of free resources of tea disease images,a total of 1375 images of five common tea diseases with similar symptoms were collected,and the disease images were expanded to 7358 by data augmentation methods.MobileNetV2 was selected as the network model of this research,and other classical convolutional neural network models were based on the tea disease dataset by making experiments,which proved that the MobileNetV2 model has obvious advantages in model size and forward inference speed.(2)In order to solve the problem that the tea disease image is easy to over-fitting during model training,and to ensure real-time recognition on the mobile,transfer learning is used to optimize the classification performance of tea disease on the MobileNetV2 model: First,MobileNetV2 is used to train the pre-trained model based on PlantVillage which is an free dataset.And then use the pre-training model to make transfer learning based on the tea disease dataset.In the meantime,the original network is optimized by these two ways: BN layer parameters are merged into the convolutional layer to simplify the structure of MobileNetV2,and the weight of the Depthwise Convolution in the depth separable convolution structure is L2 regularized to constrain parameters of the network.The result of Experiments shows that these methods have improved the accuracy of disease classification by about 6.5% and the forward inference speed up by 31% compared with training MobileNetV2 from scratch based on the tea disease dataset.(3)Transplant tea disease classification model to the mobile device,and implement a system of tea disease recognition system based on Android.The recognition result was optimized by custom camera and recognition enhancement strategy.After testing in real smartphones,the system has an average recognition accuracy rate of 95.6% for 5 types of tea diseases,and can be used normally in different types of Android devices.