Research On Quality Of Soybean Using Hyperspectral Imaging And Machine Vision Technique

It is important for quality inspection and quality classification in agricultural products processing and business. The detection technology will influence product quality, market competitiveness, and the labor intensity. In many agricultural products, soybean takes a special place in our country. It is not only the main source of vegetable oil in China, but also important source of vegetable protein. Meanwhile, soybean is one of the biggest agricultural products of China’s imports. At present, the soybean quality largely depends human eye and traditional chemical detection method.There are some insufficient aspects in objectivity, accuracy and speed by humen eye. Chemical detection will take long time, break samples and pollute environment.This research used soybean produced in Hei Longjiang province to study the rapid and nondestructive detecting methods applying hyperspectal image, machine vision, image processing, spectral analysis, pattern recognition, chemometrics and soybean science. The main research contents of this paper are as follows:(1) Soybean isoflavones content is forecasted using spectral information.Soybean isoflavones has many important physiological activity, This research uses hyperspectral image technology to detect the contents of soy isoflavones.Because of large data in hyperspectral images, principal component analysis (PCA) was applied to reduce the dimension of hyperspectral data.Using the support vector regression (SVR)、BP neural network (BPANN)、partial least squares (PLS) and stepwise multiple linear regression method (SMLR) to establish the prediction model of soybean isoflavones content.(2) Several soybean varieties were classified using hyperspectral image technology.Collecting 1000-2500 nm range of hyperspectral image data, using principal component analysis method to find out principal component images. Characteristic parameters were extracted from each feature image based on gray level co-occurrence matrix. Eight major characteristics were selected from 24 characteristic variables.Using artificial neural network to establish identification model for different soybean varieties. Soybean varieties of dongnong 4400, he feng 40, ken dou 16, hei nong 44 and he feng 25 were identified applying the identification model.(3)Using image information to identify normal soybeans, disease spot soybeans, mildew soybean, insect damage soybeans, broken soybean. The various segmentation methods were compared in this study. This study investigated morphological characteristics, color features and texture features of 39 feature parameter values, including circular degree, compactness, volume, arc ellipsoid, eccentricity, elliptic long axis and short axis and a series of characteristic parameters. To extract the characteristic parameters of data standardization, using principal component analysis to reduce data dimension, and the complexity of the neural network was reduced, and the training accuracy was improved. The BP neural network was optimized using particle swarm optimization algorithm.(4) Correlation analysis between isoflavone content and appearance quality was studied based on the average vlue of H, S, V, R, G, B and othe feature parameters.The aim of this study was to explore the feasibility of the technology to detect appearance and inner quality of soybean based on hyperspectral image and machine vision. The research results will provide principles for detection equipment of soybean, and prepare basis for on-line detection of soybean quality.The results and conclusions are as follows:(1) In view of the soy isoflavones content detection, the results show that the 1516 nm,1572 nm,1691 nm,1716 nm and 1760 nm wavelength can be used as a characteristic wavelength. After PCA dimension reduction, support vector regression (SVR) model was established. The determined coefficient R2 was 0.9713 of the model and the MSE is only 0.087.Compared with other mainstream multiple regression analysis, PCA algorithm combined with the SVR was more effective, which superior to BP neural network, PLS, SMLR and stepwise linear regression method obviously.(2) BP neural network was established to identify soybean varieties. The accuracy rate was 96% in training set, and the total predicted rate was 92.5%. Results showed that the hyperspectral image classification and recognition technology is feasible for soybean varieties.(3) Features of soybean morphological and color make the recognition result more objective, and the recognition efficiency was improved. This research captured morphology and color features using digital image processing technology. In order to extract quantitative morphology and color features, soybean grain appearance quality detection method was discussed based on morphology and color characteristics.(4) The training of the neural network belongs to the high complexity of the nonlinear optimization problem, the particle swarm optimization algorithm is a kind of new stochastic global optimization technique, and the training algorithm is competitive in neural network. It uses parallel global search strategy, the weakness of traditional BP algorithm is overcome which is easy to fall into local optimum. This paper optimized artificial neural network applying particle swarm algorithm to detect soybean appearance quality. The experimental results showed that the neural network based on particle swarm optimization algorithm improved the performance of the prediction model.(5) It is positive correlation for energy, H and G, and it is negative correlation for B, L, eccentricity ratio and circularity.However, soybean in Hei Longjiang province interbreeds, mixing harvest and mixing storage. The quality is uneven, the index of the specificity is lower than the foreign products, but the price is higher than similar imported products.Because of lacking of international competitiveness, the soybean exports is influenced seriously and the use of domestic soybean is affected in enterprises. In recent years, due to the agricultural sustainable development and the development of soybean processing products, soybean demand increased rapidly, therefore the research on soybean quality detection technology has high theoretical value and practical significance.