Research On Classification Of Crop Disease Based On Deep Learning

Crop diseases have a wide range and great harm in the process of crop planting,and are an important factor affecting the quality and yield of crops.Based on deep learning,the disease images in the actual growth environment of crops are studied,so as to classify similar diseases,which is conducive to the development of disease control and the improvement of the qual ity of agricultural products.In complex shooting scenes,crop diseases present indistinguishable similar characteristics,resulting in confusion in disease classification and low accuracy.A multi-branch fine-grained crop disease classification model TMFD-Net based on Swin-Transformer is proposed.First,in order to extract the image features of crop diseases,a Swin-Transformer combined with a hybrid attention mechanism is designed as the backbone network,and a multi-branch cascade method is used for multi-scale feature fusion to maximize the retention of disease feature information and improve the accuracy of disease image classification.Secondly,in order to be closer to the actual shooting scene,data enhancement technology is used to simulate the real environment of crop disease images.The experimental results show that the comprehensive evaluation results of the TMFD-Net on the enhanced Plant Village and Plant Pathology 2021-FGVC data sets are better than the classical crop disease classification models,and have the ability to classify similar diseases.In view of the problem that the deep learning-based crop disease classification model has a large amount of parameters and a high computational memory footprint,it cannot be effectively deployed on mobile inspection equipment.An improved lightweight crop disease classification model SIR-Net based on ResNet-34 was constructed.In order to ensure the effectiveness of basic feature extraction,a lightweight channel attention mechanism is introduced into the original residual convolution module to construct an RN-BLOCK to suppress redundant feature information.In order to further realize the model lightweight,the DGN-BLOCK is built by applying grouped convolution and depthwise separable convolution to reduce the computational complexity of the deep semantic feature extraction stage;at the same time,the light-weight channel attention mechanism is combined to strengthen the information correlation between channels and improve the ability of fine-grained disease feature extraction.The experimental results show that the classification error rate of the SIR-Net is lower than that of the ResNet-34,and the number of parameters is about 1/17 of that of the ResNet34,and it has certain deployment capabilities.Based on the TMFD-Net,a crop disease classification system is designed.The system consists of four modules,namely image management module,disease classification module,disease query module and result visualization module.By displaying the functions of the realized modules,it is proved that the system can effectively classify crop disease images.