| Food safety has always been the focus of consumers.Although the traditional pesticide residue detection method on food processing line can accurately detect pesticide residues,it cannot do online detection.Hyperspectral imaging technology and electronic nose technology are new nondestructive and on-line detection methods.Hyperspectral imaging technology can collect the image information and spectral information in the visible light-near infrared area.Electronic nose technology mimics the mammalian olfactory system,using an array of sensors to detect the composition of the gas.In this paper,the method combining hyperspectral imaging technology and electronic nose technology was used to detect apple pesticide residues.Firstly,the hyperspectral image was preprocessed.In order to solve the influence of uneven light source intensity and dark current in various bands caused by the spectral camera itself,black and white correction was carried out,and the light intensity compensation was carried out on the hyperspectral image to eliminate the influence of uneven light caused by the shape of the apple itself.Then in this study,automatic and accurate ROI selection is realized through two steps.In the first step,texture information such as contrast,sharpness and average gradient of spectral images under each band is calculated and weighted by entropy value method to obtain the optimal image segmentation band.In the second step,the Canny operator is used to extract the edge of the image,and the closing operation is carried out to connect the small holes.Finally,the position information of the sample is obtained through detection in both horizontal and vertical directions,and the background information of the image is removed,and the image is finally segmented accurately.Secondly,the spectral characteristics of the region of interest were calculated: spectral mean and spectral variance.After through the continuous projection method(SPA)method respectively characteristic bands selection,and will get spectral characteristics through the most relevant-minimum redundancy(m RMR)fusion algorithm,and builds model of multiple classification: particle swarm-support vector machine(PSO-SVM),extreme learning machine(ELM),random forests(RF),simple bayesian(NBM).By comparing multiple indicators,it can be found that the PSO-SVM model is the optimal classification model.In addition,the results in the PSO-SVM model compared with the features after direct fusion and MRMR screening were 95%.After hyperspectral feature extraction is completed,feature extraction is carried out on the electronic nose data.Five features,including stable mean,integral,maximum value,maximum wavelet energy and maximum variance of wavelet reconstruction coefficient,are extracted respectively.Variable importance projection method(VIP)is used to rank each feature value to form 50 new subsets.By selecting the first 35 VIP ranked data in the SVM model as the final electronic nose feature data,and comparing the detection accuracy of the data in PSO-SVM and Convolutional Neural Network(CNN),it is shown that the maximum VIP-PSO-SVM can reach83.33%.Finally,the hyperspectral and electronic nose feature data were fused by genetic algorithm-partial least squares(GA-PLS),and the final fusion features were obtained through GA-PLS screening,and then put into PSO-SVM,ELM,RF and other models.Through comparison,it can be found that whether direct fusion,or GA-PLS screening,The PSO-SVM model has good detection performance,and the detection accuracy of GA-PLS-PSO-SVM is up to 98.33%.To sum up,this paper used the combination of hyperspectral imaging technology and electronic nose detection technology to realize the detection of different pesticide residues on apple surface.It is proved that this model can be applied to actual food processing,and can also provide a theoretical basis for the realization of real-time nondestructive testing and classification of batch samples in agricultural automatic production line. |
PDF Full Text Request