Detection Method Of Potato Slight Bruise Based On Different Level Information Fusion Of Hyperspectral Image

As the fourth largest staple food in future, potato proved more edible value and commercial value. Whether the potato potato is available or not is directly related to quality grading,which is more related to staple potato strategy that be successfully implemented.the method of potato quality detection,can not only improve the edible value of potato, but also improve the potato industry value. Slight damage potato samples were considered as the research object, then adopt a v-shaped plane mirror hyperspectral image with information fusion method based on different levels of information fusion, and through the comparative analysis the best detecting method of the slight damage was determined. The specific research results were as follows:(1) set up a "V" flat mirror hyperspectral image acquisition platform, determined the hyperspectral image acquisition parameters for potatoes, V plane mirror hyperspectral image acquisition platform is verified by the experimental and it can be used to detect slight damage.(2) determined the data-level fusion method of potato slight damage detection. 180 slight damage potatoes and 142 normal potatoes were the research object, then used V plane mirror hyperspectral image acquisition platform for the potato hyperspectral images, in hyperspectral image 3 sub-image in F1, F2 and F3,then extracted spectral data X11, X22 and X33 spectrum matrix, proceeded the data of original data fusion layer, then used Support Vector Classification(SVC) as a modeling method, by comparing the Standard Normal Variable transformation(SNV), Multi-source Scatter Correction(MSC) and Poisson transform(Poisson Scaling) three different spectral preprocessing methods, SNV proved the most significant effect of SVC model,which was the optimal data preprocessing method for data fusion layer. Then Used the Ant Colony algorithm(ACO) characteristics of potato damage variable screening, used the nine spectral variables as input, using Fruitfly Optimization Algorithm(FOA), Genetic Algorithm(GA) and the Grid search method(Grid search) three methods for SVC penalty factor c and kernel function e in the model optimization, the FOA was the most effective, the model built by SVC for minor damage identification accuracy of calibration set and test set were both 100%, which determined the FOA for the data-layer fusion of potato slight damage detection as the optimal model parameter method.(3) determined the feature-layer fusion method of potato slight damage detection. With v-shaped plane mirror hyperspectral image acquisition platform for potato hyperspectral image, from the three sub-image F1, F2 and F3 to extract data X11, X22 and X33, adopt theSNV after pretreatment by using three kinds of unsupervised dimensionality reduction method to extract the damage features, after Support Vector Classification(SVC) model was established. The research results showed that through KPCA, SVC model had the highest recognition accuracy, and then respectively used Particle Swarm Optimization(PSO),Invasive Weed Algorithm(IWO), Bat Algorithm(BA) and FOA on the penalty factor c and kernel function parameters e of SVC to carry on the optimization, through comparative analysis, through the SVC model of PSO to optimize the optimal result, calibration set and test set recognition accuracy reached 93.48% and 93.48% respectively, thus determined KPCA as the potato slight damage detection of feature dimension reduction method of feature-layer fusion, PSO method for the optimal model optimization.(4) determined the decision-layer fusion method of potato Slight damage detection.180 slight damage and 142 normal potatoes were considered as the research object, then used v-shaped plane mirror hyperspectral image acquisition platform for the potato hyperspectral image, and from the three sub image F1, F2 and F3 X11, X22 and X33 were extracted, using KPCA to X11, X22, X33 for feature extraction, respectively. Used SVC,Extreme Learning Machine(ELM) and Random Forest(RF) to extract feature. To X11,,ELM model has the highest precision,the recognition accuracy rate reached 86.51%, the calibration set test set recognition accuracy rate reached 84.12%; To X22, RF model had the most accurate recognition accuracy rate reaching 88.84%, the calibration set test set recognition accuracy rate also reached 89.72%; To X33, SVC model had the highest precision,the recognition accuracy rate reached 89.3%, the calibration set test set recognition accuracy reached 85.98%. So ELM, RF and SVC can be potato slight damage modeling method of information fusion. By fusion criterion to X11, X22 and X33decision-layer fusion, the fusion model for identification accuracy of test set was 89.71%.(5) determined the best fusion method of potato slight damage detection. Used modeling time and model recognition accuracy as comparison basis, comparing the data-layer, feature-layer and decision-layer fusion. Through comparative analysis, the data-layer-fusion, the-feature-layer-fusion and the decision-layer fusion model of recognition accuracy were respectively 100%, 85% and 89.71%, the modeling time were respectively 3.32 s, 5.05 s and 2.42 s, data-layer fusion had the highest precision,and the modeling time was shorter. Finally data-layer fusion was used to detect the optimal information fusion method of potato slight damage.