Research On Image Classification Technology Of Apple Diseases Based On Convolutional Neural Network

In recent years,the consumption of apples is huge,and how to increase the production of apples has become a key concern of agricultural experts and fruit farmers.Studies have shown that diseases are the main factors affecting the yield and quality of apples.How to effectively identify the types of apple diseases and reasonably prevent and treat them has become a key issue that people are concerned about.In my country,the method of expert manual identification is generally used to identify apple diseases,but it is difficult to complete the efficient identification task with limited strength.Nowadays,agricultural experts have applied image technology to disease identification.Therefore,the main task of this paper is to apply image processing technology to the classification and recognition of apple disease images,optimize the network model,and improve the performance of the optimized network model on the task of apple disease image classification and recognition.The main work is as follows.The diseases of apple leaves only account for a small part of the leaves in the early stage,focusing on the diseased parts of the leaves will help improve the ability of the network model to identify diseases.Therefore,the attention mechanism is introduced into the network model structure design,and three attention mechanism modules are designed: DCA dual-channel attention module,PAM parallel mixed attention module,and SAM series mixed attention module.Taking Res Net50 as the basic network architecture,three neural network models,DCA＿Res Net50,PAM＿Res Net50 and SAM＿Res Net50,are designed by embedding the attention mechanism module.The three network models were applied to the task of apple disease image classification and recognition,and the accuracy were 96.15%,95.73%,and 96.36%,respectively.Compared with the Res Net50 network model,the recognition accuracy has been significantly improved,and the AUC indicator has also been significantly improved.It fully shows that the introduction of the attention mechanism into the design of the network model can not only effectively improve the recognition accuracy of the network model for apple diseases,but also improve the adaptation performance of the network model to the recognition task.Aiming at the small sample size of apple disease image dataset,this paper uses the idea of transfer learning to optimize the training process of the network model.In the transfer learning training,the trained Res Net50 model is used as a pre-training model by sharing parameters,and its weight parameters are transferred to the optimized network model embedded with the attention mechanism,and the network model is trained according to the set training strategy.After training by sharing parameters,the three network models DCA＿Res Net50,PAM＿Res Net50,and SAM＿Res Net50 achieved 99.02%,98.95%,and 98.25% accuracy respectively in the task of apple disease image classification and recognition.It is fully proved that the method of optimizing training using transfer learning can further improve the accuracy of the network model for apple disease classification task and improve the convergence speed of the network model.Aiming at the problem of unbalanced sample size between categories of apple disease image datasets,this paper proposes to introduce focal loss function into model training to strengthen the influence of image categories with small sample size and difficult-to-train samples on the loss function.Balancing the training error due to imbalanced sample sizes between classes.In the experiment,it is proved that the recognition accuracy of the model can be effectively improved by setting appropriate hyperparameters,and it fully shows that the optimized network model trained by the focal loss function can effectively eliminate the influence of the imbalance of sample size between the categories of the dataset and improve the model generalization performance.