Research On Peach Leaf Disease Images Recognition Based On Deep Learning

Peach is an important economic fruit tree in China.Its yield and quality affect the development of national economy.Due to the infection of leaf diseases,peach yield and quality decreased,resulting in serious economic losses.The method of peach farmers to identify peach leaf diseases is mainly based on empirical knowledge,which has strong subjectivity,high working intensity and low accuracy.In recent years,with the in-depth study of deep learning,deep learning is widely used in crop disease image recognition.Deep learning uses self-supervised learning to avoid the subjectivity problem caused by manual feature extraction.It can automatically extract the feature information of disease images,and express the abstract semantic information of disease images through high-dimensional features,so as to quickly and accurately identify disease images.In order to improve the accuracy of peach leaf disease image recognition,this paper studies convolution neural network and recurrent neural network based on deep learning.The main research contents are as follows.（1）Aiming at the problem of low recognition accuracy of peach leaf disease image caused by ignoring the location information of peach leaf disease and the context correlation information between disease regions,a disease recognition algorithm based on convolutional recurrent neural network was proposed.When convolutional neural network is used to identify peach leaf disease images,convolution operation and pooling operation are executed locally,respectively.There is a problem of missing disease location information and context correlation information between disease regions.Therefore,the bidirectional long-term and short-term memory network is introduced into the convolutional neural network to extract the bidirectional feature of the disease feature sequence,memorize the location information of the peach leaf disease and mine the context correlation information between the disease regions.The experimental results show that the recognition accuracy of the proposed algorithm on the test set is 93.73 %,which has high recognition accuracy.（2）Aiming at the long tail distribution of peach leaf disease images,a dual-channel recognition algorithm based on decoupling representation learning and classifier is proposed.Decoupling representation learning and classifier algorithm is an important method to solve the problem of long tail distribution recognition,but there are problems of poor feature representation ability and loss of head class accuracy.The proposed dual-channel recognition algorithm is based on the decoupling representation learning and classifier algorithm,which consists of two stages: representation learning and classifier learning.Image Net pre-training parameters are introduced in the representation learning stage and the parameter fine-tuning is carried out to enhance the feature representation ability.The double channels in the classifier learning stage are fixed to represent the learning parameters in the learning stage and pay attention to the head class and the middle and tail class respectively,and the final classification results are obtained through the residual fusion mechanism.The experimental results show that the overall recognition accuracy and head class recognition accuracy of the proposed dual-channel recognition algorithm on the test set are 93.81 % and 94.21 %,respectively,which effectively improve the overall recognition accuracy and head class recognition accuracy of peach leaf long tail distribution disease images.（3）Aiming at the problem of obvious intra-class differences in peach leaf disease images,a peach leaf disease image recognition algorithm based on spatial-temporal feature fusion is proposed.There is a life cycle of peach leaf disease.The color,shape and size of the lesion are different in different disease periods,and have certain temporal characteristics.Therefore,combining convolutional neural network with recurrent neural network,spatial and temporal features are fused to enhance the ability to learn effective and specific representations.At the same time,in order to enhance the feature extraction ability of different spatial positions in the image,the 3×3 convolution operation in the convolution neural network is replaced by the internal convolution operation.The experimental results show that the recognition accuracy,accuracy,recall and F1 values of the proposed algorithm on the test set are 94.87 %,93.61 %,92.54 % and 93.07 %,respectively,with good recognition effect.