The Research Of Tomato Leaf Image Classification Based On Deep Learning

Accurately judging crop diseases is an important link to monitor crop growth and condition.The development of computer vision image technology and deep learning technology provides technical conditions for precision agriculture and the possibility of intelligent agricultural development.In this paper,the tomato image disease classification as the breakthrough point,using a series of image processing technology to optimize the image data set,on this basis,two neural network algorithm to complete the tomato leaf image disease classification and disease degree classification research.The main contents of this paper include:(1)Preprocess the data.In this paper,the experimental data from Plantvillage engineering provide the tomato image data set,the first to data cleaning and filtering of the data set,the second in order to meet the requirements of neural network training,to achieve a good training effect,the number of images by simple physical means to expand,and select a series of traditional data image processing method,is used to remove the shadow performance such as the bad part of the image.This article also used in network training part of the image for further image cropping tomato disease category,through the block operation,extract only the symptoms of image center position,make the symptoms for network input data sets,the proposed algorithm experimental part of the data sets show that the symptoms to improve the positive role of network training effect.The original data set and tomato disease symptom data set constitute the tomato disease data set trained by network in this paper.(2)Combined with the transfer learning algorithm,a tomato leaf image disease classification model based on the pre-training model ResNet-34 was established.Stored in large data sets on the ImageNet trained ResNet-34 network parameters,on the original network structure,add the connection layer further optimization features,tomato disease classification used in this paper,on the basis of introducing Sample Pairing data enhancement method,improve the network classification performance,the network model can fast convergence.Part in experiments,by adjust the network effect parameter combination experiment,frozen migration parameters of different network module and verify the validity of the migration learning algorithm,finally through a network model with three common and also in migration of tomato disease classification under the background of study comparing two related models,on different evaluation index to evaluate the classification results,this paper validates the effectiveness of the algorithm,training the final network for tomato disease classification accuracy rate can reach 97.06%..(3)Combined with the hierarchical structure of tomato leaf image data,a classification model of disease degree based on multi-task learning was established based on ResNet-50 network module.Network consists of three parts: share common features of network is responsible for providing training,is responsible for tomato disease categories of coarse-grained network to classify,fine-grained network add SE module is made of three parallel network module,responsible for classifying degree of tomato diseases,including coarse-grained and fine-grained network between the cascade operation information.The network can effectively classify the tomato disease degree,and the classification accuracy reaches 93.97%.In the experimental part,the feasibility of the network structure in this paper is explained through the visual output of the feature graph in the shared network,the influence of cascading operation on the network,and the judgment of parallel network grouping.Finally,taking the classification of main task disease degree as the target,the superiority and effectiveness of the proposed network model are verified by comparing with three common flat architecture networks.