Soybean Seed Detection Based On THz Spectrum And Deep Learning

Soybean has become an important grain and oil crop in China because of its rich nutrition,variety,easy survival and high yield.Many soybean varieties with similar appearance are easy to be wrongly selected only by manual screening,which will cause huge economic losses.Therefore,the soybean market needs efficient and accurate soybean detection methods.Traditional soybean testing methods are time-consuming,laborious and highly specialized,which cannot meet the needs of ordinary practitioners in the soybean market and are difficult to popularize.In this paper,based on THz spectral technology and deep learning method,the detection methods of soybean species,component content and health status were studied:(1)Different soybean varieties and disease detection methods based on THz spectroscopy and DNN model were studied.THz time domain spectrum and THz frequency domain spectrum of 10 kinds of soybean and 4 kinds of soybean diseases were collected.SG-pca,SG-kpca,wt-pca and wt-kpca methods were used to perform the pre-processing of denoising and feature extraction for the collected THz time-frequency domain data.The DNN discriminant model of soybean varieties and disease varieties was established and compared with BPNN and SVM discriminant models.The experimental results showed that the discriminant accuracy based on THz time domain spectrum was significantly higher than that based on frequency domain spectrum.When SG-pca pretreatment method was adopted,the discriminant accuracy of DNN model of soybean varieties reached 99.18%.When the wt-kpca pretreatment method was used,the DNN model identification accuracy was 96.67%on average(2)The soybean component content detection method based on THz spectroscopy and LSTM prediction model was studied,the component contents of protein,isoflavone and water in soybean were detected.THz time-domain spectrum and THz frequency-domain spectrum of ten kinds of soybean were collected,and SG,SG-pca,WT and wt-pca methods were used to conduct denoising and feature extraction pretreatment for the collected THz time-frequency domain data.LSTM prediction models for protein content,isoflavone content and water content of soybean were established respectively,and compared with PLS and SPSS prediction models.Comparing the relative errors,correlation coefficients and root mean square errors of the prediction results of the three models,the results show that the prediction effect based on THz frequency domain spectrum is significantly higher than that based on THz time domain spectrum.When using SG pretreatment method to process time-domain data and establish LSTM prediction model,the prediction result of soybean component content is the best.The relative error,correlation coefficient and root-mean-square error of the prediction results of protein,isoflavone,water content in the first stage and water content in the second stage were 0.0359,0.9968 and 0.0205;0.0318,0.9978,0.0098;0.0386,1.0000,0.0044;0.2934,1.0000,0.0055;(3)Study on soybean seed health detection method based on CNN model.The normal and diseased soybean picture sets were collected through the micro-focus lens,and the CNN detection model of soybean health status image was established and compared with BPNN and SVM image detection model.The experimental results showed that the CNN model was significantly more effective than the contrast model in detecting the soybean health status.