Research On Food Subsurface Quality Detection Technology Based On Raman Scattering Image

Spectral detection technology(Hyperspectral,Near-Infrared,Raman Spectroscopy)has been widely used in food quality inspection,which is limited by the depth of photon penetration and can only be used for non-destructive detection of the surface layer of food.The sealed packaging and the outer skin/shell of the food can effectively guarantee the stable quality of the food during storage,transportation,and marketing.Meanwhile,the surface layer(packaging,outer skin/shell,etc.)also blocks the optical signal transmission reflecting the internal quality,and the interaction reaction between the surface material itself and the excitation light also generates interference spectra,which will affect the accurate estimation of the internal quality of food products,and there is no non-destructive detection technique for the subsurface layer of layered foods that applies to different surface layer disturbances.To address such issues,this study takes food subsurface quality detection as the research object,acquires Raman scattering images using a line-scan-based Raman hyperspectral detection equipment,identifies,separates,and enhances subsurface Raman signals based on spatially offset Raman spectroscopy(SORS)combined with signal processing techniques,and finally qualitatively and quantitatively assesses the quality of subsurface food products through a machine learning modeling analysis method.To further evaluate the applicability of the SORS data analysis method to the field of food subsurface detection,an improved Monte Carlo simulation is used to generate optical characteristic models commonly used in the food field,and the performance and optimization strategies of the data analysis method under different characteristic parameters are investigated,and finally,the transformation of the point source-line source structure in the Raman spectroscopy detection equipment is guided to enhance the SORS technology in practical applications.The specific contents are as follows:(1)The variation of scattering profile at different spatial offset distances in Raman scattering images contains the organizational spatial distribution information of stratified samples.The traditional subsurface detection method based on the spectrum at the optimal offset distance leads to the problem of spatial physical information loss and is easily affected by optical properties and instrument noise.Using Raman spectra at a specific offset distance will lead to the loss of spatial physical information.Raman spectra and scattering feature extraction technology based on an Attention-based LSTM network is proposed.The scattering profile of each band is used as input,and the physical distribution and chemical composition information of the tissue to be measured is extracted by the LSTM module with the time-series characteristics of spectral data.The output of each module is weighted by the attention mechanism,and the robustness of the prediction model is further improved by combining the fusion features of fully connected networks.The freshness detection of shelled shrimp meat by an Attention-based LSTM network combined with SORS technology proves that this technology avoids human error when selecting offset distance,effectively utilizes the physical information of tissue distribution of samples to be tested,realizes high-precision freshness evaluation of whole shrimp with shells,and also provides a candidate detection scheme for internal food quality detection with surface interference.(2)To address the problems of surface layer signal interference left in the feature extraction of Raman scattering images of layered samples and detection difficulties in adapting to diverse layered substances,a subsurface feature peak identification method based on a generalized Gaussian model is proposed based on the consistency of the scattering profile of each layer of substances can be characterized by the parameters of the fitted curve,which combines the molecular specificity of Raman spectra to achieve the detection of subsurface substances and components.Firstly,the scattering image is processed to extract typical Raman spectral peaks to ensure the main information required for the detection process,and then the attenuation profile of the spectral peaks of offset laser points is standardized to ensure the consistency of the scattering characteristics of each layer of material,and the spectral peak attenuation profile is fitted by a generalized Gaussian model to reduce the dimensionality of the data,and finally,the fitted parameters of the generalized Gaussian model are input to the hierarchical clustering model and a suitable threshold is selected to identify the subsurface material Raman spectral peaks to reconstruct the Raman spectra.The technology was applied to the simulated internal food inspection with packaging,and the applicability of the model to packaging thickness,surface,and subsurface materials was verified.Finally,the detection of white granulated sugar in packaging shows that the model can effectively identify and separate the Raman characteristic peaks produced by packaging and food,eliminate the interference of surface spectral peaks,and realize complete nondestructive detection of packaged food.(3)Aiming at the problems of weak spectral peaks,overlapping spectral peaks,and baseline background information loss in the subsurface Raman feature peak recognition method,an improved fast independent component analysis algorithm is proposed to separate the full Raman spectra of subsurface samples,to ensure the fidelity of subsurface spectra and rich information mining.Firstly,the movable quadratic mean of information entropy is used to select the observation feature matrix of the Raman scattering image to reduce redundant information,and two main independent components are separated by the fast independent component analysis method,then independent component attribution recognition is carried out by using the attenuation characteristics of SORS characteristic peak signal,and finally,non-negative separated signals are ensured by baseline recognition and correction.The technology is applied to the signal separation of food with packaging,and the applicability of this method was evaluated by three simulated packages and four simulated internal food materials.Finally,three practical packaged foods were used to prove that this method can effectively separate the subsurface Raman spectrum signals and retain the information of weak peaks,overlapping peaks,and baselines,which can be used as a pretreatment method and an auxiliary analysis method for packaging food detection.(4)Existing food subsurface analysis methods based on spatial offset Raman spectroscopy are usually limited to specific sample material backgrounds,and it is difficult to fully evaluate the applicability of SORS analysis methods and explain the optical transmission theory of SORS.Therefore,a subsurface analysis method based on statistical replication Monte Carlo simulation(SRMC)and spatial offset Raman spectroscopy is proposed to generate a complete mechanism model with food optical characteristics to evaluate the algorithm goodness of the spatial offset Raman spectroscopy data analysis method.Firstly,the photon density distribution generated by two-step SRMC is used as the weight of each voxel.Then,each voxel is used as an excitation source to generate a Raman spectrum and detect the photon contribution of each point on the scanning line.The weight and the photon contribution convolution of this point constitute Raman scattering image.SRMC technique is used to generate 5625 sets of Raman scattering images with different optical characteristic parameters to evaluate signal processing methods.Experiments show the influence of six optical characteristic parameters on Raman scattering,and the advantages of the Fast ICA method in analysis speed,automation level,and analysis accuracy,as well as adaptability under different parameters.(5)Given the low signal-noise of the Raman offset spectrum obtained by line scanning detection equipment based on point laser,Long-term exposure increases cumulant,and multipoint collection takes a long time,and it is difficult to realize batch detection and pixel-level visual distribution.A point-line conversion technology based on Monte Carlo simulation and SORS technology is proposed to detect food subsurface.Firstly,two-step statistical replication Monte Carlo simulation is used to generate the spectral contribution of line scanning imaging system based on point laser source and line scanning imaging system based on line laser source,and the transformation models of point laser equipment and line laser equipment are obtained.Then determine the optimal offset distance of line scanning equipment based on the line laser source,adjust the line light source equipment according to laser power,and verify the effectiveness of this technology by quantitative adulteration detection of packaged sugar.The point-line conversion technology based on Monte Carlo simulation can be extended to the ring collection and ring light source detection equipment corresponding to SORS,which can improve the batch detection ability of packaged food and adapt to diversified samples.