| Tea has important economic and living value,but there are dozens of common tea diseases,which seriously affect the yield and quality of tea.Therefore,accurate identification of tea diseases is the prerequisite and guarantee for tea planting and disease prevention.The current convolutional neural network has been widely used in plant disease recognition,but there are problems such as insufficient data set and high morphological similarity,which leads to unsatisfactory recognition accuracy.In view of this,this article carried out a series of identification studies on the original data sets of 8 common tea leaf diseases,hoping to obtain effective improvement methods to further improve the accuracy of tea disease identification.The main research contents and results are as follows:1)Tea disease identification based on improved Res Net network.First,compare the current six mainstream optimizers to find an optimizer suitable for the tea data set.Then,network modification was carried out for Res Net.The convolution layer with convolution kernel size of 7*7 was replaced by 3 convolution layers of 3*3,and a maximum pooling layer was added in front of the full connection layer for dimension reduction.Finally combine the modified Res Net model with a suitable optimizer,and use particle swarm optimization algorithm to optimize the initial learning rate in the optimizer.The results show that using this method to train the tea disease model,the test accuracy rate reaches to 89.47%,while the test accuracy rate of the common Res Net network model at this stage is only80.45%.2)Tea disease identification based on transfer learning.A training method of secondary transfer learning is proposed.Firstly,the Res Net network is pre-trained with Image Net data set.Then the pre-training model was trained by using the plant disease data set.Finally,the tea disease data set was used to train the model after transfer learning.The results show that the recognition accuracy is increased from 89.47% to 93.98% after one migration,and the recognition accuracy is increased to 96.24% after the second migration training.3)Validation of different models based on secondary migration.In order to verify the universality of the secondary migration method and compare different tea disease identification models,the VGGNet model and Google Net model were first tested for unmigrated and secondary migration,and then the Res Net model,VGGNet model and Google Net model after secondary migration were compared.The results show that the accuracy of VGGNet and Google Net secondary migration test are improved,which are 87.63%and 97.74%,respectively.Among the three secondary transfer models,Googlenet model has little difference with Res Net in the recognition effect of tea disease,while VGGNet has the worst recognition effect. |
