A Fine Identification Model For Multi Category Pests And Diseases

Pest identification plays an important role in maintaining plant safety,understanding crop plant safety,crop yield,and timely,effective and reasonable treatment.The traditional identification of pests and diseases is usually manual identification or machine learning.Relatively few applications of deep learning technology are used.Most of them use deep learning to identify different diseases of a certain type of crops or different types of crops.The same disease or a few categories of crops.Pests or diseases,lack of identification for large sample sizes.With the continuous development of deep learning,the convolutional neural network as a very important part has also developed.Convolutional neural network is a kind of deep neural network.It has its unique features that can reduce the number of parameters during training,and has special convolutional layers that other deep neural networks do not have.Because of its different connection methods,the method is a local connection method,and the weight value can be shared,its downsampling layer will further improve the robustness of the network,and it can also reduce the complexity of the network and facilitate operation.At the same time,the overfitting problem of some networks has also been greatly improved.This paper starts with the basic concepts and algorithms of convolutional neural networks,deeply studies the theory and related problems of convolutional neural networks,and finally proposes the PD-Net(Pest and Plant Diease Net)network,which can effectively solve complex and large-scale problems.The diseases and pests of the samples are difficult to identify,and other evaluation indicators such as the accuracy of identification and loss value have been improved to a certain extent.The main work of this paper is as follows:(1)First,we used two multi-category large-sample disease datasets and insect pest datasets,and then self-collected a portion of the disease datasets with a small sample size,namely,general cherry powdery mildew,severe cherry powdery mildew,and grapevine.General black rot,severe grape black rot,general tomato leaf blight,severe tomato leaf blight,general citrus yellowing disease,and severe citrus greening disease,100 images for each type,a total of 800 images,and then these images were integrated into the disease classification recognition in the dataset.(2)Secondly,this paper proposes a deep learning-based pest and disease identification network model architecture PD-Net,which incorporates an attention mechanism and a cross-layer non-local module,which solves the problem of the large sample size in the data set and the need for key features.Difficulty in extraction and representation(blurred images and complex backgrounds)and unclear representation of finely differentiated features(large differences in the same type,small differences in different types).(3)Finally,we conducted a large number of experiments,first of all,we conducted an ablation experiment,and tested the original model with the improved points and improved models proposed in this paper.The ablation experiments proved that the two improved methods we added are real and effective.,and then carry out comparative experiments to compare the network model proposed in this paper with Alex Net,VGG16,Google Net,Inception-v3 and Dense Net121 models,and conduct experiments on disease data sets and insect pest data sets respectively.For verification,according to the final experimental results,it can be concluded that under the same environment as the learning rate,the number of iterations,the weight decay rate,and the loss function,the accuracy rate,recall rate,F1,and loss value of the network model proposed in this paper are all A certain improvement,and at the same time,the effectiveness of the algorithm model in this paper has been further verified.In summary,this paper adopts two large-scale disease and insect pest data sets,and collects some disease data sets by itself,and proposes corresponding improvement methods and new network model architecture for the problem of pest identification.Through ablation experiments and comparisons Experiments verify the effectiveness of the proposed network architecture.At the same time,it will also provide new methods and ideas for other scholars who study pests and other identification and classification problems in the future,which also makes it possible to solve the problem of classification of large-scale pests and diseases in the past.Convenient.