Studies On Classification Of Foodborne Pathogens Based On Hyperspectral Imaging Technology

Food-borne diseases caused by food-borne pathogens seriously affect the quality of people’s life and health.Therefore,it is of great significance to develop a rapid,accurate,low-cost and wide-ranging detection technology for food-borne pathogen detection.Taking Escherichia coli,Staphylococcus aureus and Salmonella as research objects,this thesis combined hyperspectral technology with chemometrics to study the ability of hyperspectral technology in detecting bacterial colonies cultured for different culture periods and on different media.The main research results were as follows:(1)The optimal methods for bacterial colony classification and detection under different culture time based on hyperspectral technology were determined and the optimal bacterial culture time was identified.In this study,hyperspectral images were acquired on bacterial colonies cultured with three time gradients of 24±2h,48±2h and 72±2h.Morphological processing was used to determine the region of iterest of bacterial colonies and extract bacterial spectral data.Different preprocessing methods and different wavelength selection methods were used for the three time gradient data,and the full wavelength and characteristic band models based on partial least squares discriminant analysis,support vector classification optimized by genetic algorithm(GA-SVC)and optimized by particle swarm optimization(PSO-SVC)were established respectively.The results showed that the PLS-DA model could only classify bacterial samples cultured for 48±2h,while the SVC model could detect all samples from any culture time.By comparising the bacterial detection cycle and detection accuracy,it was determined that 24±2h was the best culture time for bacterial classification.Moreover,the optimal full-wavelength model was the GA-SVC model(All-SG-GA-SVC)established based on all of the SG smoothed spectra of the three incubation time samples.The corresponding classification accuracy rates of calibration set and prediction set and Kappa coefficients were 98.69%,98.75% and 0.981,respectively.The best simplified model was the GA-SVC simplified model(All-SG-GA-GA-SVC)established after the SG spectral smoothing and GA wavelength selection of the overall based on the data of samples from the three culture time,and the corresponding classification accuracy rates of calibration set and prediction set and Kappa coefficients were 98.69%,98.57% and 0.979,respectively.(2)The method for classifying bacteria cultured on different culture media based on hyperspectral technology was determined.In this study,hyperspectral images of bacterial colonies cultured on three different general media(Luria-Bertani agar LA,Plate count agar PA and Tryptone soy agar TSA)for 24±2h were collected and the average and pixel-level spectra of bacterial colonies were extracted.PLS-DA and SVC full wavelength and characteristic wavelength bacterial classification models at colony level and pixel level were established,respectively.For the colony-level classification of bacteria,only the SVC full-wavelength model and the simplified model could realize the detection of bacteria on different culture media.The best full-wavelength model was the GOA-SVC model established after MSC spectral preprocessing(MSC-GOA-SVC),where the corresponding classification accuracy rates of calibration set and prediction set and Kappa coefficients were 99.45%,98.82% and 0.982,respectively.The best simplified model was the GOA-SVC model established after the data was processed by MSC and by CARS wavelength selection(MSC-CARS-GOA-SVC)and the corresponding classification accuracy rates of calibration set and prediction set and Kappa coefficients were 99.45%,98.73% and 0.980,respectively.For the pixel-level bacteria classification,a convolutional neural network with Le Net-5 network structure could achieve accurate and stable prediction of bacteria.When the number of iterations was 1000 and the size of the convolution kernel was 1×13,the classification accuracy of the model could be as high as 97.07%.This study provided a simpler and more convenient detection method for the detection of pathogenic bacteria cultured on different media and layed a foundation for the development of portable bacterial detection instruments.