Research On Feature Fusion And Recognition Of Potato Typical Disease And Insect Pest Images

As a kind of important grain crop,potatoes are usually attacked by diseases and insect pests.Depending on the experiential distinguish of agricultural workers while observing crop leaves,the traditional recognition method for diseases and insect pests is considered to be detecting restrictively,labor exhaustive and running slowly,which can easily cause reduction of potato production and quality.With the development of computer technology,image processing and pattern recognition theory were used to establish gradually the visual recognition system,which has been a forward trend in the direction of agricultural intelligence.The image of potato diseases and insect pests under natural conditions,were studied as research objects in this recognition system.Automatically,potato diseases and insect pests were classified based on image processing and pattern recognition technology.The processing steps from images acquisition to images recognition for diseases and insect pests were as follows.Firstly,based on characteristics for the images of diseases and insect pests,I channel image of HSI color model could be filtered by Median filter algorithm and R,G,B channel images of RGB color model were segmented roughly by Grabcut algorithm.As thus,by using two-dimensional Otsu method which could segment finely a channel images of Lab color model,and morphological algorithm which was easy to deal with binary space image,the clear and complete targets of diseases and insect pests were separated from natural background.Secondly,in accordance with the invariable visual information of diseases and insect pests,a new texture feature extraction method,High frequency covariance matrix eigenvalues and low frequency lower order moments in wavelet domain(HELM),was proposed.Effectively improving target portrayed level,this method could be considered as a function to overcome the target problem,such as translation,rotation,zooming and so on.Thirdly,on the basis of 435 dimensional extracted features of color,shape and texture,a new adaptive algorithm,Feature weighted fusion based on principal component analysis(FWFPCA),was proposed.Quickly eliminating the high dimensional feature disaster,the algorithm could be used for enhancement of feature expression ability.Finally,in view of the fusion features for 13 types of diseases and insect pests,decision tree thought was adopted to train Support vector machine(SVM)classifiers step by step,which had a underlying advantage to achieve the purpose of sample prediction.Based on MATLAB R2012 b platform,features recognition experiments were carried out on potato diseases and insect pests,3 types of diseases and 10 types of insect pests,respectively.The main experimental contents and results were as follows:(1)Portrayed level of HELM textural features was proved.Texture features in space domain were traditionally related to LBP and features based on Gray-level co-occurrence matrix(GLCM).Texture features in wavelet domain were traditionally related to LBP of wavelet domain(LBPW),Low frequency lower order moments(LM),and High frequency lower order moments and low frequency lower order moments(HMLM).Compared with SVM recognition rates of LBP,GLCM,LBPW,LM and HMLM,it was found that proposed HELM features could not only increase recognition rate,but greatly reduce feature dimension.(2)Feature expressive ability of FWFPCA algorithm was verified.Feature fusion algorithms were usually related to Straight ranking selection(SRS),Feature block ranking selection(FBRS),Straight principal component analysis(SPCA),Feature block principal component analysis(FBPCA),and so on.Compared with SVM recognition rates of SRS,FBRS,SPCA and FBPCA,it was found that proposed FWFPCA had a higher recognition rate.(3)Recognition performance of SVM classifier was testified.Pattern classifiers were mainly related to Artificial neural network(ANN)for which three layers BP network structure was constructed,and Bayes classifier to which Parzen window function was applied.By using adaptive FWFPCA fusion features,average recognition rates of proposed SVM increased at least 2.10%,4.10% and 5.20% than ANN and Bayes,and run times of proposed SVM were 2.33 s,3.22 s and 1.06 s faster than ANN.So,By proposed FWFPCA algorithm and SVM classifiers,proposed HELM features combined with optimal color and shape features,will not only ensure effectively the recognition accuracy but improve greatly the recognition speed.