Research On Detection And Classification Of Crop Diseases Based On Convolutional Neural Network

Agriculture is the basic industry that has supported the construction and development of China’s national economy for many years.Various crop products brought by agriculture are important resources to ensure the normal operation of people’s life.Research on the rapid and accurate identification and diagnosis methods of crop diseases has important practical significance for the timely prevention and control of crop diseases,reducing economic losses,reducing environmental pollution,and ensuring the quality and safety of agricultural products.In order to implement the rapid detection and classification of crop disease,this paper studies the detection and classification of crop disease leaves.The main research contents of this paper are as follows:(1)Aiming at the problem of the limited ability of traditional feature recognition,a fusion method of depth features extracted by convolutional neural network and traditional methods was proposed,and random forest feature screening method was used to screen the fused features,and finally support vector machine was used to classify the filtered features.Tested in Plant Village dataset,94.08% classification accuracy was obtained.(2)Aiming at the problems of high similarity,different sizes,and difficulty in locating crop disease leaves,a detection algorithm based on an improved feature pyramid and attention mechanism was proposed.Based on the ResNet50 feature extraction network,this algorithm combines PANET’s feature aggregation pyramid and Libra R-CNN’s feature balance pyramid to obtain a new improved feature pyramid and uses both channel attention mechanism and spatial attention mechanism to strengthen the feature pyramid.The experimental results show that the tomato disease leaf detection algorithm based on improved feature pyramid and attention mechanism has the best detection performance on the data set in this paper,and effectively improves the detection accuracy of small targets.(3)Aiming at the difficulty of image classification of crop diseases under complex natural background,a leaf classification algorithm of apple diseases was proposed based on multi-scale feature fusion and channel attention mechanism.Based on ResNet50,the single-scale convolution operation in the residual module is replaced by the pyramid convolution that can simultaneously extract multi-scale features,and the number of parameters is reduced by using the dilated convolution kernel instead of the large convolution kernel.The multi-scale features are optimized by using the channel attention mechanism to obtain more expressive features.The experimental results show that the classification accuracy of apple disease leaves can reach 94.96% by combining multi-scale features and channel attention mechanism.(4)A web-based crop disease image detection platform was designed.Through the deep learning algorithm deployed in the server,remote detection of disease images taken and uploaded on site is realized by using Web pages.