| Grapes are susceptible to leaf diseases in the growth process,leading the decline of yield and quality,which brings severe economic losses.Therefore,effective recognition and prevention of grape leaf diseases are of great importance for the healthy growth of grapes.Deep learning and computer vision provide a more effective solution for monitoring and diagnosing crop disease images.Early grape leaf disease images mainly rely on manual collection and marking,which requires a lot of time and cost to obtain the diseased image samples,and the types and quantity of pieces are limited.Most image recognition algorithms based on deep neural networks require large training datasets.When training image samples with only a few labels or even no labels,the model is prone to problems such as overfitting,which reduces recognition accuracy and robustness.To solve the above problems,based on deep learning method,this paper took grape leaf black rot,ring spot disease and brown spot disease as the research object,and studied how to improve the identification accuracy of grape leaf disease in small samples.The work of this paper includes:1.A dataset enhancement method was proposed based on a variational autoencoder and a generative adversarial network.The application of generative adversarial network GAN and variational autoencoder VAE in unsupervised data enhancement was explored.Combined with the structural characteristics of GAN and VAE to generate data by learning data distribution,a data enhancement model based on variational autoencoder countermeasure generation network(VAAGN)was constructed by integrating image visual information and semantic information.The VAAGN data enhancement model with strong generalization ability was pretrained on the PlantVillage dataset,and the VAAGN model generated the diseased images to enhance the few-shot grape leaf disease dataset,which provided support for improving the accuracy and robustness of the disease recognition model.2.A few-shot grape leaf disease recognition method based on improved CNN was proposed.The structural characteristics and training strategy of convolutional neural network were analyzed.Combined with the data distribution of grape leaf disease dataset,the VGG-16 model structure was improved to construct the grape leaf disease recognition model(GV).With the transfer learning method,the weight of the VGG-16 model pretrained on Image Net dataset was transferred to small sample grape leaf disease recognition.Besides,the impact of training methods(new learning and transfer learning),whether the dataset was enhanced,and model parameters on the performance of GV model was explored to improve the training and recognition efficiency of GV model on grape leaf disease dataset.3.A few-shot grape leaf disease recognition method was proposed based on improved CNN and Transformer.The application of CNN and Transformer model image recognition was studied.The convolution layer of pretrained VGG-16 model was used for local feature extraction.The high-level features and long-term dependencies were obtained through Transformer encoder.Besides,the VGG-16-Transformer(VTF)model was constructed to recognize grape leaf disease.Analyze the training and recognition effect of VGG-16 model,transformer encoder model,and VTF model on the enhanced grape leaf disease data set,and evaluate the VTF model,to provide a reference for improving the recognition accuracy of grape leaf disease in small samples. |
