| As the main food crop in China,wheat plays an important role in agricultural production,circulation and consumption in China.The unsound kernel of wheat seriously affects the quality and food security of wheat.At present,the automatic detection methods of grain quality mainly include volatile matter detection method,image analysis method and terahertz detection method.In view of the hysteresis of volatile matter detection method in imperfect grains of wheat,image analysis method and terahertz detection method are vulnerable to the influence of the environment.High spectral analysis method can achieve rapid nondestructive detection,which is a potential detection method.In order to detect wheat perfect grains and six kinds of unsound kernel wheat,this paper proposes a method of unsound kernel wheat recognition based on hyperspectral imaging.The main research contents are as follows :Obtain hyperspectral data of unsound kernel wheat.A total of 300 samples of 7 kinds of wheat grains in each category were prepared,and the hyperspectral imaging system of imperfect wheat grains was constructed.The visible-near infrared hyperspectral data of 2100 wheat grains were collected,and the threshold segmentation,filtering and other preprocessing were carried out to determine the region of interest of wheat grains.The Savitzky-Golay algorithm was used to analyze the hyperspectral image of wheat grains,and 261-dimensional effective hyperspectral data at 420.61-980.43 nm were selected.Identification of unsound kernel wheat by machine learning.SPA algorithm was used to reduce the dimension of 261-dimensional hyperspectral data,and 33-dimensional hyperspectral data such as 420.61 nm and 503.06 nm were selected.PCA algorithm was used to extract the optimal spectral wavelength,and the optimal spectral wavelength was 647.57 nm.Eleven morphological features such as density and rectangularity of wheat grain image were extracted,and 12 texture features such as energy and entropy were extracted.The feature space was established by the combination of ‘ spectrum ’,‘ morphology + texture ’,and ‘ spectrum +morphology + texture ’.Each type of wheat grain was trained by 250 hyperspectral data,and50 grains were used as the test.The PSO-SVM model was established to classify imperfect grains,and the recognition rates were 93.17 %,87.14 % and 97.14 %,respectively.The recognition rate of ’ spectrum + shape + texture ’ is the highest,which is 10 % higher than that of ’ shape + texture ’.The ACO algorithm is used to analyze the 56-dimensional features of ’spectrum + shape + texture ’,and 26-dimensional features such as 422.64 nm are selected.Compared with the PSO-SVM model before feature optimization,the recognition rate is increased by 0.33 %.Identification of unsound kernel wheat by deep learning.The images of 647.57 nm,568.36 nm and 591.78 nm were selected by PCA algorithm to establish the data set.There were750 training sets for each type of wheat grain image and 150 verification sets.The wheat CNN network and Mobile Net V2 network are constructed.The accuracy of training set and verification set of Mobile Net V2 network is 99.92% and 97.88%.Compared with CNN,Mobile Net V2 network improves accuracy by 1.05% in training set and 2.66% in validation set.Compared with PSO-SVM model and Mobile Net V2 network model,the recognition rate of Mobile Net V2 network model is improved by 0.74%.Through the machine learning method,it is necessary to extract image features to establish feature space,and deep learning can extract features through the network itself.Thus,the deep learning method is more suitable for the identification of wheat imperfect grains,and it is feasible to identify unsound kernel of wheat by hyperspectral imaging. |
