Research On Crop Leaf Disease Detection Method Based On Convolutional Neural Networ

Crop diseases are an important factor leading to a decrease in crop yield and threatening China’s food security.Accurately identifying crop diseases is an important measure to ensure crop yield.In the actual agricultural production process,growers often lack sufficient knowledge of plant pathology,making it difficult to accurately identify crop diseases.In recent years,automatic disease recognition technology based on deep learning has received widespread attention.However,there is a problem with small intra-class differences and large inter-class differences in crop leaf diseases,and the detection accuracy and efficiency in traditional deep learning models are not high.To address these issues,this article improves existing convolutional neural networks and designs specific lightweight neural networks to improve the effectiveness of crop disease detection.The specific research content is composed of the following parts:(1)A lightweight improved Mobile Net V2 model,CA-Mobilenet V2(Coordinate Attention),was proposed to accelerate model convergence while improving detection accuracy.The main improvements are as follows: 1)The lightweight coordinate attention module is embedded in the Mobile Net V2,which improves the accuracy and reduces the extra computing cost.2)Add Tanh Exp activation function for lightweight network to accelerate the training and convergence of the model,and enhance the robustness and generalization of the model.(2)As the complexity of diseases and real-world interference increases,the optimization space of traditional convolutional neural network models reaches a bottleneck,and the network’s feature extraction ability and robustness are challenged.In response to these issues,a lightweight crop disease recognition model called DFCANet is proposed.DFCANet is mainly composed of two parts: the Double Fusion with Coordinate Attention(DFCA)module and the Down-Sample(DS)module.The DFCA module includes double feature fusion and coordinate attention modules.To fully fuse shallow and deep features,different hierarchical features are doubly fused,improving the model’s feature extraction ability.The CA module can suppress background noise and focus on diseased areas,improving the model’s robustness in complex environments.In addition,the DS module is used for down sampling.It reduces information loss by expanding the feature channel dimension and using deep convolution.Comparing with classical convolutional neural networks and excellent algorithms in domestic and foreign literature,the results show that DFCANet has higher accuracy in identifying crop diseases.Additionally,the DFCANet model has parameters and Flops of 1.91 M and 309.1M,respectively,which is lower than traditional deep convolutional network models and most lightweight network models.(3)Crop leaf disease identification system.The model is deployed to the mobile APP,so that the model has a good visual application and disease warning effect.