Detection Of Internal And External Defects Of Potatoes Based Onsemi-transmittance Hyperspectral Imaging Technology

A strategy was launched to make potatoes to be staple food in China, which will promote the development of potato industry greatly, and also reflect the huge edible and commercial values of potatoes. As an important part of potato industry, quality detection plays a significant role in the process of the strategy propulsion. Therefore, the development of nondestructive detection methods for potato quality control is of great scientific significance and good application prospects.In this thesis, internal defect(hollow heart) and external defects(green rind and bud) potato samples were selected as research object. A semi-transmission hyperspectral image acquisition system was constructed to acquire the hyperspectral images of the potato samples and the single defect detection models and multi-defect detection models of internal and external defects were built by using semi-transmission hyperspectral imaging and data processing methods. The results were as follows:(1) A semi-transmission hyperspectral image acquisition system was developed to acquire the hyperspectral images and the feasibility of this acquisition system was verified by experiments.(2) The semi-transmission hyperspectral imaging method for detecting potatoes of hollow heart was determined. 149 normal samples and 75 hollow heart samples were selected as research object, based on which support vector machine(SVM) was utilized to develop detection models. 5 preprocessing methods were used to preprocess the original spectra, and by comparing the different influence, normalize was determined to be the optimal spectral preprocessing method. Competitive adaptive reweighed sampling algorithm(CARS), successive projection algorithm(SPA) and CARS-SPA secondary selection method were utilized to select important variables from the 520 wavelengths respectively. By comparing accuracy of model which was based on the selected variables, CARS-SPA was found to be the optimal variable selection method ultimately, and 8 variables were selected. Artificial fish swarm algorithm(AFSA), genetic algorithm(GA) and particle swarm optimization(PSO) was used to optimize the SVM model parameters: penalty parameter c and kernel parameter g. The different SVM models which used the 8 selected variables as input were built based on the 3 parameter optimization algorithms. By comparing accuracy of the models, AFSA was found to be the optimal parameter optimization method ultimately. Recognition rate of test set and calibration set of the corresponding optimal model could both reach 100%.(3) The semi-transmission hyperspectral imaging method for detecting potatoes of green rinds and buds was determined. 3 kinds of modeling methods, including partial least squares discriminant analysis(PLSDA), K nearest neighbor(KNN), adaptive boosting(Ada Boost), were used to build detection model for the randomly placed green rind and bud potatoes, and the optimal preprocessing method of each model was confirmed. The results show that, for the detection of green rind potatoes, the recognition rate of 100% for the test set and calibration set could be obtained by the PLSDA model preprocessed by the SNV and the Ada Boost model preprocessed by mean center or multiplicative scattering correction(MSC). For the detection of budded potatoes, 100% of the recognition rate of test set and calibration set could be obtained by the PLSDA model preprocessed by MSC and the Ada Boost model preprocessed by Mean Center or SNV.(4) The semi-transmission hyperspectral imaging method for detecting internal and external defect potatoes was determined. Normal, budded, green-rinded and hollow-hearted potatoes were used selected as research object, and the 4 targets classification models were built by SVM, KNN and error correcting output codes least squares support vector machine(ECOC-LSSVM). By comparing the 3 models, ECOC-LSSVM model was confirmed to be the optimal multi-class classification model for the detection of internal and external defects in potatoes. In order to simplify the model, 4 kinds of manifold learning methods, including locally linear embedding(LLE), supervised locally linear embedding(SLLE), isometric mapping(Isomap) and kernel principal component analysis(KPCA), were used to reduce the dimensionality of spectra. Based on the low dimensional spectra data, 4 kinds of ECOC-LSSVM models were built. The results show that, SLLE was comfirmed to be the optimal dimensionality reduction methods, and single recognition rate of normal, bud, green rind and hollow heart potatoes was 96.83%, 86.96%, 86.96% and 95%, respectively. The hybrid recognition rate was 93.02%.(5) The multi-class classification detection model for internal and external defects potatoes was optimized. Because green-rinded and budded potatoes are both inedible, this two class external defects potatoes were combined as one class, which was used to build three target classification model with normal and hollow heart potatoes. SLLE and ECOC-LSSVM were used to reduce the spectra dimensionality and build classification models. The single recognition rate of normal, green rind and bud combination class, hollow heart was 98.41%, 100%, 95%, respectively. The hybrid recognition rate of three class was 98.45%.