Research On Tea Plant Disease Identification Method Based On Improved DenseNet

Tea is easily infected by diseases and insect pests in the growth process,which seriously affects the quality and yield of tea.Therefore,only by accurately identifying diseases can we carry out effective control.With the continuous development of machine learning and artificial intelligence,convolutional neural network and other technologies have been successfully applied to the automatic identification of plant diseases and pests.However,with the deepening of the number of network layers,the data labels required for model training are also greatly increased.At the same time,the amount of calculation of network model is also larger and larger,and the requirements for operating equipment are higher and higher,which limits the popularization and application of deep learning algorithm in the field of crop disease detection.Considering that the amount of data that may exist in the image acquisition of tea plant disease is too small and the uneven distribution of different kinds of diseases will affect the accuracy and robustness of the depth neural network model,this paper studies the recognition of five common tea plant disease.It is expected to design effective recognition methods to improve the recognition accuracy of tea plant disease in the scene of small sample data set and uneven sample distribution.The main research contents and results are as follows:(1)In view of the problem that the small number and uneven distribution of samples in tea plant disease data set will affect the accuracy of deep learning model,a SE-DenseNet-FL model based on transfer learning is proposed.Firstly,based on DenseNet and SENet,this model proposes a network model SE-DenseNet,which integrates the advantages of both.SENet can increase the local receptive field by integrating the channel attention mechanism to improve the recognition effect of the network model;Then,the focus loss function is introduced to replace the loss function in the original DenseNet to obtain the SE-DenseNet-FL network model to alleviate the imbalance of sample categories;Finally,based on the characteristics of transfer learning,first pre train the model on the large public data set plantvillage,migrate the pre training model to the enhanced tea plant disease data set,and fine tune the parameters to improve the accuracy and robustness of the network model in the case of small samples.Finally,the recognition and classification accuracy of SE-DenseNet-FL model for tea plant disease reached 92.66%,Experiments verify the effectiveness of SE-DenseNet-FL network model in the scenario of small samples and uneven sample distribution.(2)In view of the large amount of parameters and calculation of the proposed SEDenseNet-FL network model,a lightweight model SE-DSDenseNet-FL based on dual channel depth separable convolution is proposed.Firstly,the model combines the of dense intra block composite functions in DenseNet 3×3 convolution is replaced by deep separable convolution to reduce the parameters of traditional convolution;Then after the bottleneck layer 3×3 deep separable convolution is processed in parallel,and the outputs of the two channels are added,so as to further reduce the model parameters.The final tea leaf disease identification model only costs very little,but greatly reduces the amount of parameters and calculation of the model.By applying the improved model to the tea plant disease identification system,it is verified that the model can better meet the needs of disease identification in the actual tea garden and has a good application prospect.