Research On Crop Disease Image Recognition Method Based On Model Fusion And Distillation

China is a world agricultural powerhouse,and crop pests and diseases are the main cause of economic losses in the country’s agriculture.Therefore,how to effectively identify and prevent pests and diseases has become a current research hotspot and challenge.In recent years,with the rapid development of the field of computer vision,the performance of crop disease recognition based on deep learning has been greatly improved.However,current research mainly focuses on the detection performance of single models,which is prone to model preferences.Additionally,there are still problems with model generalization for crop disease recognition tasks in different scenarios.Furthermore,reducing the computation and complexity of the model and improving its operational efficiency are also necessary conditions for the deployment of intelligent agricultural machinery and equipment platforms.Therefore,this article focuses on the research of the model generalization and computational efficiency problems that currently exist in the research,including:(1)A crop disease recognition method based on model fusion is proposed to address the defect that a single neural network model may have learning biases towards training data.Firstly,four mainstream convolutional neural networks,ResNet50,DenseNet121,Xception,and MobileNetV2,are evaluated for their single model performance.Then,feature-level and decision-level model fusion are applied to each of these four models,and the recognition results are outputted.The feature-level fusion method applies averaging,maximum and concatenation compression fusion to the last output feature layer of each sub-network,achieving efficient complementarity of heterogeneous features.The decision-level fusion method applies maximum and averaging fusion to the output probabilities of each sub-network,achieving efficient joint probability distribution decision-making.Experimental results show that the feature-level fusion method outperforms the decision-level fusion and single-model methods.The concatenation compression feature fusion method achieves the highest recognition accuracy,reaching 98.44% and90.36% respectively in the Simplify-PDR2018 dataset with a simple background and the CCR dataset with a complex background.Furthermore,the cross-library experimental results on the Plant Doc data subset and actual captured images also indicate that the feature fusion method has better accuracy and generalization performance than the single-model method.(2)A mixed knowledge distillation algorithm is proposed to address problems such as the large number of model parameters and high computational resource consumption.Traditional knowledge distillation methods rely on teacher models,but in the human learning process,a strategy of teacher teaching and student reviewing is often used to achieve self-learning.Inspired by this,this paper uses a mixed learning paradigm of bidirectional knowledge distillation and self-distillation to imitate human learning mechanisms.For bidirectional knowledge distillation,a bidirectional knowledge distillation framework is constructed between the teacher and student models,in which prediction information is exchanged between them to complete the cycle of knowledge transfer,achieve complementary heterogeneous feature information,and obtain the best-performing student model.For self-knowledge distillation,the student model outputs its own soft label distribution based on its own weights,and the real labels of the student model are smoothed to consolidate its own knowledge.Two mixed knowledge distillation strategies are obtained by adding self-distillation inside and outside the bidirectional distillation framework.DenseNet201 and ResNet152 are used as teacher models,and MobileNetV2 and ShuffleNetV2 are used as student models.Experimental results show that the mixed knowledge strategy with self-distillation applied outside the bidirectional distillation performs the best,achieving recognition accuracies of 73.17% and 88.15% on the CIFAR100 dataset and the PDR2018 crop disease dataset,respectively.