Research And Application Of Deep Learning In Crop Disease Image Recognition

Crops growing in natural environment are vulnerable to multiple diseases,which seriously affect their yield and quality.Therefore,timely identification of crop disease types and taking targeted measures are of great significance to ensure the healthy growth of crops.Traditional disease identification methods are costly and inefficient,and can not meet the requirements of modern agriculture.With the continuous development of artificial intelligence,the use of deep learning technology for crop disease image recognition has become a research hotspot.Convolutional neural network as a representative network can automatically extract features for end-to-end learning and obtain better recognition results.However,there are also problems such as high computing costs and difficult deployment at mobile terminals.In order to improve the recognition accuracy and training speed of the model,and reduce the calculation cost and scale of the model,this paper carries out the research on crop disease image recognition based on depth learning.The main contents are as follows:1.For apple leaf disease images,a recognition method based on mixed attention mechanism residual network(CBAM-Res Net)and Transfer learning is proposed.By improving the original residual block of Res Net50 and embedding the concatenated channel attention module(CAM)and spatial attention module(SAM),the model can pay more attention to the important information area in the training process,and effectively extract the features of apple leaf lesions;At the same time,Transfer learning is used to redesign the full connection layer of the original model and train it,which improves the expression ability of the network,and the recognition accuracy can reach 95.5%.2.For tomato leaf disease images,an improved lightweight model(MCA-Mobile Net)and a recognition method based on Wasserstein distance Generative adversarial network(WGAN)are proposed.The improved multi-scale feature fusion module and Coordinate attention module(CA)are embedded on the basis of Mobile Net V2,so that the model can improve the recognition accuracy and reduce the calculation cost;At the same time,WGAN was used to enhance the original data set,which enriched the tomato disease image data and improved the robustness of the model.The recognition accuracy can reach 94.11%,and the parameter quantity is only 2.75 M.3.For grape leaf disease images,a recognition method based on improved lightweight model(HC-Efficient Net)and Ensemble learning is proposed.Based on Efficient Net V2 and combined with mixed cavity convolution(HDC),the model is improved to further enlarge the receptive field and obtain more disease features;The coordinate attention module(CA)is introduced to replace the SE module of the original model,which improves the global attention ability of the model;At the same time,the integrated learning of multiple classifiers is used to improve the recognition effect of the model.The recognition accuracy can reach95.29%,and the parameter quantity is only 2.06 M.This paper studies different data sets from two perspectives of improving model performance and reducing model size,providing a new method for in-depth learning in crop disease identification.