Research On Crop Disease Identification Based On Lightweight Convolutional Neural Network

Agricultural production is the foundation of human survival and the cornerstone of national economic development and security.How to ensure agricultural production safety and increase crop yields is a topic of widespread concern among scholars in many fields around the world.The combination of artificial intelligence and agriculture is an important means to promote agricultural digitization,intelligence,and modernization,as well as an important way to improve agricultural production efficiency and competitiveness.Currently,crop disease identification based on convolutional neural networks has achieved high accuracy,but its parameter volume is large and requires a certain amount of computing resources.When faced with a crop production environment with limited computing resources,it cannot provide services.Although providing technical services through cloud deployment can solve this problem,it also faces problems such as slow response speed and high application costs.In view of the limited computing resources in crop production environments,this paper optimizes the design of lightweight network structures based on research on lightweight neural network models,further improves network performance,and constructs a lightweight network model with better comprehensive performance:(1)Lightweight crop disease identification model construction method.A lightweight convolutional neural network model,Mobile Net-Rep MLP,is proposed for crop disease identification scenarios with limited device computing resources.The network greatly reduces the overall parameter volume of the model by reducing the number of channels of pointwise convolution and introduces the Rep MLP module to extract rich crop disease features even in low-dimensional space.Experimental results on the Plant Village dataset show that the Mobile Net-Rep MLP model achieves high recognition accuracy and greatly reduces the number of parameters,which is 59% less than the original model.(2)Construction of crop disease identification models in complex scenarios.For crop disease identification scenarios with complex backgrounds,Mobile Net-Rep MLP enhances the model’s ability to extract crop disease features from complex backgrounds by introducing the ECA attention mechanism module,further improving the model’s expression ability.The knowledge distillation algorithm is also used to allow the model to learn information about disease features that were previously ignored,thus improving the generalization ability and accuracy of the model.These two methods can effectively improve model performance without significantly increasing the number of model parameters and computational effort.The experimental results show that introducing the ECA module can improve model performance and make the model perform well in this scenario.After the knowledge distillation algorithm,our model’s recognition accuracy is further improved,reaching the highest accuracy among compared lightweight network models while having the least number of parameters.In the model inference speed experiment,our model still performs well.Compared with the original Mobile Net V2 model,the inference time of a single image is reduced by 24.7%.The excellent performance of our model provides a reference for crop disease identification scenarios and application deployment.