Research On Recognition Of Specific Crop Diseases Based On Deep Learning

Accurate identification of crop diseases is of great significance to increase crop yields.Traditional crop disease identification methods mainly rely on humans,but manual identification may cause misjudgments.Methods based on machine learning cannot automatically extract disease features.Therefore,in order to realize the intelligent recognition of crop diseases,this paper studies the identification method of specific crop diseases based on deep learning method.The main research contents and results are as follows:(1)Aiming at the problem of inconsistent image size and unbalanced data categories in the specific crop disease data set,the crop disease data set was preprocessed.For the problem of unbalanced data categories,data enhancement was performed on samples with a small number of categories.For the problem of inconsistent image size,the images were compressed in batches to keep the main disease information of the image while keeping the overall quantity unchanged.In order to speed up the network convergence during training,the images were normalized before being input to the network.(2)Aiming at the problems of network over-fitting and gradient disappearance,the residual network was selected as the basic model of this paper and the Res Net50 was fine-tuned.Firstly,the fully connected layer in the Res Net50 structure was improved and its output was set to the number of sample categories in the data set in this article.Then the activation function was improved to solve the problem that some neurons cannot be activated.Finally,the Adam optimization algorithm and Focal loss function were combined to optimize the model.The experimental results show that the overall recognition accuracy of fine-tuning Res Net50 is 1.88% higher than that of Res Net50 in the recognition of single crop fine-grained diseases.And the overall recognition accuracy of fine-tuning Res Net50 is 1.19% higher than that of Res Net50 in the recognition of multiple crops fine-grained diseases.(3)Aiming at the problem of poor recognition of certain categories in multiple crops,an improved Res Net50 model was proposed.Firstly,in order to reduce the number of parameters and speed up the recognition time,three 3×3 convolution kernels were used instead of 7×7 convolution kernel.Then,to retain all the information of input feature maps and transmit important information to the subsequent structure,the paper added the max-pooling operation to the down-sampling module in the network structure.Finally,the attention mechanism was added to the Bottleneck block to effectively use the relationship between the channels of the features.The results show that the overall recognition accuracy of the improved Res Net50 is 91.63%,which is better than traditional deep learning models such as Inception v3.