Research And Application Of Crop Disease Identification Based On Deep Learning

Crop diseases are the main culprits affecting the yield and quality of modern agriculture,hence preventing and controlling the spread of diseases is an important measure to reduce economic losses.With the change of agricultural production mode and the development and maturity of image data acquisition technology,the application of crop disease diagnosis methods based on deep learning is promising.However,the existing convolutional neural network models are rarely used to diagnose crop diseases and there are still some problems such as insufficient feature extraction and utilization and low identification accuracy.Therefore,this dissertation conducts research on crop disease identification based on deep learning and applies the research results.The main work is as follows:(1)To solve the problem of low utilization of feature key information,an improved SE-Net(ISE-Net)module is proposed.In addition,a lightweight convolutional neural network model based on Alex Net,namely IAlex Net,is proposed to achieve effective identification of apple leaf diseases,in which multiple 3×3 convolutional kernels are used instead of large convolutional kernels in order to reduce network parameters,average pooling is used instead of flattening,and the convolutional output channels are improved while Dropout is discarded.The comparison experiments with Vgg16,Alex Net and Mobile Net＿V2 models show that the IAlex Net model has the highest accuracy of 95.2% for apple leaf disease identification,which verifies the effectiveness of the IAlex Net model for apple leaf disease identification.(2)To address the problems of inadequate feature extraction and poor model effectiveness when the data is unbalanced,a multi-feature fusion module that uses multiple convolutional kernels to extract features in proportion to each other is proposed.In addition,the ISE-Net module is introduced to distinguish the correlation and importance between features,and then a crop leaf disease identification model named IDense Net based on the multi-feature fusion module and Dense Net is proposed.By comparing with Res Net34,Vgg16,Dense Net,Goog Le Net and Mobile Net＿V2 models,the experimental results show that the IDense Net model reaches the highest accuracy of 96.1% for the unbalanced dataset.The ablation experiments show that,in the multi-feature fusion module,the convolutional kernels of sizes 1×1,3×3 and 5×5 have the strongest ability to extract features in a ratio of 1:2:1,which indicate that the IDense Net model is also fasible and effective when dealing with unbalanced data.(3)In view of the few application of deep learning methods used in the Huanglong disease identification,the proposed IAlex Net and IDense Net models are applied to identify the Huanglong disease.The comparative experiments with Vgg16,Mobile Net＿V2 and Alex Net models show that the IAlex Net model has the highest accuracy of 94.8%.Moreover,in the comparison experiments of Res Net34,Dense Net,Goog Le Net and Mobile Net＿V2 models,the IDense Net model achieves the highest accuracy of 96.1%.These verify that the two new models can effectively identify the Huanglong disease,but the deeper IDense Net model outperforms the IAlex Net model.