Image Recognition Of Four Kinds Of Fruit Tree Diseases

As fruit trees are apt to be affected by global climate and environmental conditions,coupled with the susceptibility of fruit trees in the growth cycle,diseases may outburst,which heavily endanger the agricultural industry.Therefore,artificial intelligence technology plays great impact on identifying fruit tree diseases in a timely and accurate manner to ensure safe agricultural industry.At present,many studies have combined image processing and traditional machine learning methods to achieve higher application value.However,there still exist some shortcomings.For example,most of the studies on crop leaf image recognition use localization strategies to remove background,segment images,extract color information,texture and shape features,few studies focus on global features such as disease spot quantity;In terms of disease image model construction,the existing studies compare a variety of single models,choose one single model based with the best score.As a result,ensemble learning methods are seldom used;In terms of the effect of disease recognition,traditional recognition methods are not as effective as deep learning under the premise of insufficient feature extraction and expression.Based on the above problems,this thesis focused on pear black spot,pear rust,apple mosaic,and apple rust that have high incidence in feruit trees,and studied the leaf feature extraction methods of the four diseases.Using the strategy of combining traditional machine learning algorithms with ensemble learning,we conducted the training and test of effective models with low cost on professional dataset.As a result,we got the model with equivalent even better than deep learning models and other classical models.Finally,we implemented an online disease image recognition prototype system to provide technical support and reference to identification of fruit tree disease.The main research work is summarized as follows:(1)Due to the diversity of the morphological characteristics of diseased spots,we started from the analysis of the global characteristics of the diseased leaves to select the colour features,texture features and the number of diseased spots as original features,and carried out feature extraction work.Experiments showed that the extracted features can describe the information of diseased leaves more completely and meet the needs of disease identification.(2)After the original feature extraction,this thesis standardized the data of 18 features to narrow the scale gap,and used the data dimensionality reduction method to obtain the main features to prepare the data for further high-quality model training.(3)We explored the representative classical machine learning algorithms including K nearest neighbor,logistic regression,random forest,support vector machine and Adaboost.We analyzed and discussed the recognition performance of the models on different diseases.We took the above 5 models as the base learner to build the stacking ensemble learning model.The experimental results showed that the ensemble model has significantly improved the recognition rate of all four diseases compared with the single model,and the overall average recognition rate of the model reaches 97.33%,which is better than the traditional ensemble model based on Majority strategy,and also better than the Resnet-50 deep learning model.This indicates that the advantages of deep learning are difficult to be fully realized when the dataset size is not large enough,and the effect of the Stacking ensemble learning strategy is more desirable.(4)Based on the above research,this thesis developed a prototype system for image recognition of fruit tree diseases and demonstrated the application method of the online recognition system.