| Potatoes are important food crop in the world which have large-scale cultivation throughout the world. The quality of potatoes will affect its economic value. Nondestructive Testing to potato quality, it will increase potatoes economic value. So it will has important scientific significance and good prospects to rapidly test internal and external quality of potatoes with nondestructive inspection.This paper take the sprouting, green rind, black heart and normal potatoes as the study samples which named Kexin Number 1 and the hyperspectral imaging acquisition system had been built successfully. With the integrated use of the hyperspectral imaging technology, data analysis techniques and information fusion techniques, this paper study on the identification methods to external quality of randomly placed potatoes, the identification methods to inernal quality of potatoes and the identification methods to external and internal quality of randomly placed potatoes. Comparing the recognition accuracy of potatoes’ external quality using image features in different directions, the optimal detection model to the external quality of potatoes have been determined under the using of hyperspectral image information. Comparing the recognition accuracy of potatoes’ internal quality using different spectral feature extraction methods, the optimal detection model to the internal quality of potatoes have been determined with the using of spectral information. Comparing the recognition accuracy of potatoes’ internal and external quality using different manifold learning reduction algorithms, the best detection model to the internal and external quality of potatoes have been determined with the fusion of hyperspectral imaging and spectral information.The results are as follows:1) With building the hyperspectral data acquisition platform, the sprouting, green rind, black heart and normal potato samples had been taken and hyperspectral data had been collected. To simulate the actual production, the potato defection were not deliberately sited on the camera when collecting the hyperspectral data of sprouting potatoes and green rind potatoes. The potatoes with external defection were placed in three different orientations(facing on the camera, broadsiding on the camera, facing back camera).The every acquisition mode have one third.2)Study on the recognition accuracy of potatoes’ external quality using hyperspectral image dimension.Determining the optimal image color space is the RGB color space for potato samples. The image feature parameters have been taken at 0°, 45 ° and 90 °gray level co-occurrence matrix of the potato randomly placed. Determining the division of potato sample set as random method,support vector machine and extreme learning machine model were established to determine the potato quality. The model of highest recognition rate was the combination wih 45 °gray level co-occurrence matrix and support vector machine with the mixed recognition rate of test set reached 91.13%. The single recognition rate of the sprouting,green rind and normal potatoes were respectively 87.10%, 95.56% and 89.58%.3)Study on the recognition accuracy of black heart potatoes’ quality using hyperspectral information.Determining the division of black heart and normal potato sample set were random method. Using four different pretreatment methods to processing the original spectra, autoscale had been determined as the optimal spectral preprocessing methods.Using successive projections algorithm(SPA), competitive adaptive weighting algorithm(CARS) and uninformative variable elimination(UVE) to prefer the original spectral wavelength. 20 features variables had been selected by successive projections algorithm. C ARS 431 spectral wavelength were reduced to 34-dimensional spectral data by competitive adaptive weighting algorithm. 314-dimensional features variables had been selected by uninformative variable elimination. Using successive projections algorithm to select 314-dimensional spectral variable s, eight spectral variants were got successfully.Comparison the recognition accuracy of partial least squares discriminant analysis(PLSDA), support vector machine(SVMDA) and K nearest neighbor(KNN) model, CARS-PLSDA, CARS-KNN, CARS-SVMDA, SPA-PLSDA, SPA-KNN, SPA-SVMDA, UVE-PLSDA, UVE-KNN, UVE-SVMDA and other models were established to determine the black heart potatoes. The optimum recognition model was UVE-SPA-SVMDA when detecting black heart potato samples. The mixed recognition rate reached 100% in calibration and testing set of this model.4) Study on the recognition accuracy of internal and external quality of randomly placed potatoes with the fusion of spectral information and hyperspectral image.Using none, autoscale, standardized normal variate and detrend pretreatment methods to processing the original spectra, detrend had been determined as the optimal spectral preprocessing methods.Utilizing the diffusion maps, locally linear embedding and hessian locally linear embedding to cutting down the dimension of spectrum data of potato samples after the spectral preprocessing methods which named as detrend, Six models were respectively established which were diffusion maps and support vector machine(DM-SVM), locally linear embedding and support vector machine(LLE-SVM), hessian locally linear embedding and support vector machine(HLLE-SVM), diffusion maps and extreme learning machine(DM-ELM), locally linear embedding and extreme learning machine(LLE-ELM), hessian locally linear embedding and extreme learning machine(HLLE-ELM). Comparing and analyzing the results of the 6 models, It turnet out that the diffusion maps(DM) was the best manifold learning dimension reduction algorithm to deduce the potatoes’ spectral information.Gray level co-occurrence matrix(GLCM) was determined as the image texture characteristic extracting method. Every hyperspectral image of potatoes was morphologically processed before extracting 84 image texture characteristics based on gray level co-occurrence matrix(GLCM). By using successive projections algorithm(SPA), 10 texture features were properly selected, which concentrated on homogeneity, energy, correlation,contrast.With the fusion of the spectral data and image features, Six models were respectively established by using SVM and ELM to detecting the external and internal quality of randomly placed potatoes. Comparing the results of 6 models and the spending time of DM-SVM, LLE-SVM, HLLE-SVM, DM-ELM, LLE-ELM, HLLE-ELM models, the diffusion maps and extreme learning machine(DM-ELM) model was the best model. The DM- ELM model’s mixed recognition rate for detecting the external and internal quality of randomly placed potatoes reached 96.58% and it only cost 0.11 s. The single recognition rate of test set for sprouting potatoes, green rind potatoes, blackheart potatoes and normal potatoes respectively reached 97.30%, 93.55%, 94.44% and 100%. |
