Research On Salmon Quality Detection Methods Based On Spectroscopy

Salmon is smooth,delicious,nutritious and popular with consumers.However,there are food safety issues such as counterfeiting and spoilage in the current salmon market,which have an adverse effect on the health and interests of consumers.Traditional physicochemical methods are complicated,time-consuming,costly and destructive.Therefore,it is of great significance to find out a simple、fast and accurate salmon detection method.In the paper,a comprehensive study was carried out with the application of spectroscopy,deep learning,image processing and other related algorithms and technologies,and finally the salmon detection models for salmon counterfeiting,adulteration and quality were established.The research contents are as follows:（1）Salmon species such as rainbow trout,Heilongjiang salmon,Norwegian salmon and Chilean salmon were identified using infrared spectroscopy.An outliers exclusive method based on principal component analysis-linear discriminant analysis-mahalanobis distance（PCA-LDA-MD）was proposed in this paper.The advantages and disadvantages of different pretreatment methods were compared.The spectral differences of different salmons were analyzed and an accurate spectral identification model was established.The results showed that the performance of the model was the best when the peak area normalization method was combined with PLS-DA.The determination coefficients of the calibration sets and the cross validation sets were at 0.97 and 0.95,respectively;the RMSEC and RMSECV were0.37 and 0.52,respectively.The model could significantly distinguish four kinds of fish and the prediction accuracy of test sets was 96%.（2）Adulterated or non-adulterated salmon was identified using infrared spectroscopy.The adaptability and range of different pretreatment methods were studied.The attribution of different spectral functional groups was analyzed.The intrinsic relationship between spectral characteristic peaks and differences in fat and protein composition of salmon was studied.The effect of different spectral bands on modeling was studied and an accurate spectral identification model was established.The results showed that the PLS-DA model combined with the original spectrum could accurately identify whether Norway salmon was adulterated or not,but it was difficult to identify the specific adulteration ratio.When the normalized pretreatment method was combined with the characteristic spectrum of450-1900 cm^-1,the accuracy of the model could reach 90%.The determination coefficients of the calibration sets and the cross validation sets were 0.99 and 0.98,respectively;the RMSEC and RMSECV were 2.3 and 4,respectively.（3）The salmon storage time,total viable count（TVC）and total volatile base nitrogen（TVB-N）were detected using VIS/NIR technology.The effects of salmon meat and salmon skin on the detection results were studied.The feedback of meat differences on spectral peaks was analyzed.A stack denoising sparse coding（SDAE）modeling method based on deep learning was proposed.The results showed that salmon meat had better detection results than salmon skin.The SDAE algorithm improved the denoising capability and detection accuracy of BP neural networks.When storage time was detected,the determination coefficient of test sets reached 0.98,and the RMSEP reached 0.93;When TVC was detected,the determination coefficient of test sets reached 0.96 and the RMSEP reached 0.28;When TVB-N was predicted,the determination coefficient of test sets reached 0.96,and the RMSEP reached 1.25.（4）Salmon color was classified by machine vision technology,and the freshness of salmon was detected by fusion of image and spectral information.A salmon image acquisition system based on machine vision was established.The L*a*b*color measurement algorithm based on error compensation was studied.A salmon color grading system was established.The detection method of salmon freshness based on convolutional neural network was studied.The detection method of salmon freshness based on transfer learning was studied.A novel detection method of salmon freshness by fusion of spectrum and image information was studied.The results showed that the visual system could accurately detect the color and color level of salmon.When the self-built 11-layer convolution neural network was used,the prediction accuracy of the calibration sets and test sets were 74.2%and 41.1%,respectively.When transfer learning method based on GoogLeNet was used,the prediction accuracy of the calibration sets and test sets were61.72%and 58%,respectively.When image and spectral information were used to detect salmon freshness,the model performance was significantly improved,and the prediction accuracy was increased to 93.4%.The research achievements of this paper can solve the complex,time-consuming and costly technological status of the current meat quality detection field.It can not only accurately detect the existing salmon counterfeiting and quality problems,but also provide an efficient and low-cost detection idea for other fish and livestock.