Research On Small Sample Plant Disease Identification Based On Deep Convolutional Generative Adversarial Networks

As far as some plant diseases are concerned,it’s hard to obtain a large number of samples in the experiment because of their strong infectivity,wide transmission routes and great harmfulness,and it’s difficult to achieve ideal results in the application of Deep Learning method for classification and recognition.Deep Convolutional Generative Adversarial Networks(DCGAN)is applied in this paper to expand the sample sets.So,the ratio of positive and negative sample sets is approximately 1:1 after the experimental expansion and the size of the sample sets has reached ten thousand approximately.Then the Convolutional Neural Network is applied in the expanded sample sets to classify the cankers.Finally,the Incremental Learning method is introduced to update the trained network model.For DCGAN,the paper mainly investigates the improvement of CNN.Replacing Dropout Regularization with BN and adding Mute layer behind each convolution to omit a certain number of signals(or units)without distinction during training and validation,which eliminates the effect of variance deviation to some extent.Besides,a method is proposed for evaluating the effect of generated samples from the perspective of the similarity between the generated samples and the original samples,and a better generation model is selected.Because over 80% neurons are concentrated in the fully connected layer,the speed of network training is very slow.The improvement of Alexnet model is mainly based on accelerating the training of network,which reduces the parameters of the network by an order of magnitude,improves the classification accuracy of network,eliminates the impact of sample imbalance and improves the classification accuracy by using the method of transfer learning.Finally,the method of Incremental Learning is proposed to update the network model;adding new sample sets to the network continuously to improve the generalization performance and classification accuracy of network.After Incremental Learning,the newly added data sets can also have an ideal classification result while improving the classification accuracy of the original test samples.According to the sample quality evaluation method generated in this paper,a better generation model is selected.After optimization,the training speed of Alexnet is reduced to half of the original.After adding the Incremental Learning,it’s recall ratio has increased to 97.52%,the F1 to 89.88% and the accuracy rate to 98.91%.In view of the above results,it has satisfied the need of practical application initially and laid a foundation for the automatic diagnosis,promotion,popularization and convenience of use of citrus canker disease.