Research On Detection And Recognition Methods Of Apple Disease Based On Hyperspectral Imaging

Apple is one of the most common fruits in fruit market in our country,but it is vulnerable to disease in the process of growth and stored,which causes a lot of economic losses.So it is necessary that selecting the disease apples in the early period of the fruit sorting.The detection of apple diseases mainly detection by manual sorting,but it is difficult to achieve the consistency of the classification which the difficult sorting and poor accuracy.In this research Hyperspectral imaging was used to achieve the fast,non-destructive detection for the disease apple,which had a great significance in improving the quality of the apple detecting and grading.The key achievements are contained in this research as follows:(1)Hanfu apples that were planted in large area of north China were as the research object,Though a mass of surveys,it was found that the common disease have anthracnose,brown spot disease,bitter pit diease and black fruit rot.The characteristic wavelengths were selected by three methods,which were Manifold distance,Mahalanobis distance and successive projections algorithm(SPA),in order to extract a small amount characteristics of wavelengths for detecting disease apple.It was discovered by comparison that three characteristic wavelengths(681?867 and 942nm)selected by twice successive projection algorithm(SPA2)had optimal effects in detecting disease apple.(2)In this paper,the feature vectors were selected from the texture feature in area-of-interest of the normal and diseases apples or the spectral characteristic of three spectrum relative reflectivity of characteristic wavelengths extracted by SPA2.The models of linear discriminant analysis(LDA),BP artificial neural network(BP)and support vector machine(SVM)were respectively established for detecting disease apple.Then it shows that SPA2-BP was the best detection method in detecting disease apple and the detection rate of the training set could reach 100%and validation set was 98%.Results show that a small amount of spectral information can be effectively identified disease of apples adopting the BP neural network.(3)Three different characteristic vectors were combined by the images texture characteristics that selected through the region of interest in this paper and three spectrum relative reflectivity characteristic wavelengths adopted by SPA2.The models of BP artificial neural network(BP)and support vector machine(SVM)were respectively established by these characteristic vectors combination for detecting disease apple.Then it shows that SVM detection model combined with spectral features and texture feature as the input vector was the best detection method in detecting differents disease apple.The detection accuracy of the normal fruit in validation set was 95%.The detection accuracy of anthracnose was slightly poor as 90%,bitter pit's was 95%,brown blotch's was 95%and black rot's was 95%.The test results show that support vector machine(SVM)detection model can effectively classify apples disease for detection.So that it provides theoretical basis in the multispectral on-line detection of fruit quality grading.