Research On Tomato Disease Identification Based On Convolutional Neural Network

Accurate and fast tomato picture detection is the basis of tomato disease diagnosis and prevention,and is also an important guarantee for tomato yield and quality.Traditional machine learning recognition methods require professionals to select the features of disease images manually.The quality of the selected features directly affects the model’s recognition results.and there are considerable difficulties in network migration.Secondly,the manually selected features cannot describe the category differences between the same crop’s different diseases degrees accurately,which leads to the problem of semantic understanding.The method of disease detection and recognition is relatively single and there is no research on fine-grained tomato diseases.So,this paper introduces convolutional neural networks and transfer learning into finegrained tomato disease recognition.The main research contents and innovations are:(1)Aiming at the semantic understanding bias problem of manually selected features,this paper proposed a tomato disease recognition method based on convolutional neural network.On the one hand,this paper builds three different convolutional neural networks,and discusses how the two different data augmentation methods impact the tomato disease recognition results.On the other hand,explore how different mixup ratios effect the accuracy of tomato disease identification.The experiment proves that the convolutional neural network trained by the data set which has been enhanced by the mixup data is better for tomato disease classification,and the mixup ratio of 0.4 is more suitable for model training.(2)Aiming at the idea of using transfer learning parameters initiation network to improve the accuracy of tomato disease identification,this paper studies the tomato disease identification method is based on transfer learning.In this paper,four single convolutional neural networks,Inception-v3,Xception,Dense Net,and Res Net,are used to transfer and learn tomato diseases.By fine-tuning the parameters and using the focal loss function to optimize the network,the risk of network model degradation caused by data imbalance is alleviated,and the accuracy of tomato disease identification is further improved.(3)The single model has insufficient feature extraction for fine-grained tomato diseases,so,this paper discusses the research of tomato disease recognition based on feature fusion.In order to make full use of the semantic information of different network models,this paper selects four different networks,and fusing features extracted from different networks to obtain different fusion models.On this basis,integrated the network fusion model with the optimal single model,which further improves the accuracy of tomato disease identification.This paper builds a tomato disease recognition system based on multi-model integration,which lays a good foundation for the research on tomato diseases in the future.