Near-infrared hyperspectral imaging method for real-time detection of bacterial colony count of salmon fillets

Abstract

The invention discloses a near-infrared hyperspectral imaging method for real-time detection of the number of microbial colonies of salmon fillets, and the method comprises the following steps: S1, preparing salmon samples exposed in air at different times, and carrying out the detection of the number of microbial colonies; s2, carrying out spectral imaging measurement, and collecting hyperspectral imaging HSI data; s3, performing region-of-interest identification and spectral data extraction based on the data acquired in the S2; s4, dividing samples into a training set, a prediction set and an independent external test set; s5, establishing an APC spectrum quantitative analysis model of the salmon fillets and a spectrum identification model for APC standard exceeding judgment of the salmon fillets; s6, carrying out standard normal variable transformation SNV and Norris derivative filtering NDF preprocessing on the spectral data; s7, optimizing a wavelength combination through a two-stage wavelength selection method of moving a window and gradually eliminating multiple wavelengths; and S8, verifying the precision of the model by using an independent test set. Compared with the prior art, the near-infrared hyperspectral imaging method for real-time detection of the microbial colony count of the salmon fillets has the advantages that a few-wavelength combination model is provided, and a solution is provided for developing a small-sized special NIR-HSI detector for salmon fillet detection through the near-infrared hyperspectral imaging method for real-time detection of the microbial colony count of the salmon fillets.

Description

Technical Field

[0001] This invention relates to the field of spectral technology for food safety testing, specifically to a near-infrared hyperspectral imaging method for real-time detection of microbial colony counts in salmon fillets. Background Technology

[0002] Raw salmon slices are nutritious and delicious, and are a popular food with a large sales volume. However, they are also a food that is very easy to spoil because salmon slices are exposed to air during preparation and consumption, which can cause microorganisms to grow on the surface and cause food safety issues.

[0003] Traditional fish microbial colony total (APC) detection mainly uses the plate count method, which requires reagents, aseptic operation and microbial culture. The method is complex and time-consuming, and is not suitable for on-site real-time detection of salmon slices.

[0004] Near-infrared (NIR) spectroscopy offers the advantages of rapid and convenient detection. Meat food molecules contain a large number of hydrogen-containing groups (XH), which exhibit absorption characteristics in the NIR spectral region.

[0005] When salmon fillet samples are exposed to air, they become contaminated by environmental microorganisms. The metabolic processes of these microorganisms produce a variety of new metabolites containing hydrogen groups, resulting in variations in the overall spectral patterns before and after contamination.

[0006] Therefore, NIR spectral pattern recognition has a molecular mechanism for detecting APC in salmon slices.

[0007] Hyperspectral imaging (HSI) technology combines spectral detection and image analysis, offering the advantages of non-destructive and non-contact detection, and has demonstrated its application potential in many fields.

[0008] According to literature review, near-infrared hyperspectral imaging (NIR-HSI) has not yet been applied to the real-time detection of microbial colony counts in salmon fillets.

[0009] This invention will introduce a modeling and wavelength selection method for real-time detection of microbial colony counts in salmon slices based on NIR-HSI. Summary of the Invention

[0010] The technical problem to be solved by this invention is to overcome the above-mentioned technical defects and provide a near-infrared hyperspectral imaging method for real-time detection of microbial colonies in salmon slices, which provides a solution for the development of a small-scale dedicated NIR-HSI detector for salmon slice detection using a few-wavelength combination model.

[0011] To solve the above-mentioned technical problems, the technical solution provided by the present invention is: a near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices, comprising the following steps: S1: Prepare salmon samples that have been exposed to air for different periods of time and test for microbial colony count; S2: Perform spectral imaging measurements and acquire hyperspectral imaging HSI data; S3: Based on the data collected in S2, identify the region of interest and extract spectral data; S4: Divide the samples into a training set, a prediction set, and an independent external test set; S5: Establish a quantitative spectral analysis model for APC in salmon fillets and a spectral identification model for judging APC exceeding the standard in salmon fillets; S6: Perform standard normal variable transformation (SNV) and Norris derivative filtering (NDF) preprocessing on the spectral data; S7: Optimize wavelength combinations through a two-stage wavelength selection method involving moving windows and multi-wavelength phase-out; S8: Use an independent test set to verify the model accuracy.

[0012] Preferably, step S1 includes exposing salmon slices to air for 0-24 hours, taking 4 slices per hour, and determining the total bacterial count (APC) using the plate count method.

[0013] Preferably, in step S2, a near-infrared hyperspectral imager with a spectral range of 883~2492nm is used; Four samples were placed separately on a white background paper and placed on a moving stage. The stage was moved at 10 mm / s to collect HSI data.

[0014] Preferably, step S3 includes constructing a pseudo-color image recognition salmon slice region in the NIR double spectral region and extracting the average spectrum within multiple regions of interest (ROIs).

[0015] Preferably, in step S5, partial least squares is used as the basic algorithm to establish a quantitative analysis model of salmon fillet APC spectrometry; Partial least squares-discriminant analysis is used as the primitive algorithm for the classifier.

[0016] Preferably, step S6 includes preprocessing using Standard Normal Variable Transform (SNV) and Norris Derivative Filter (NDF), and selecting preprocessing parameters based on the modeling effects of the two models.

[0017] Preferably, the moving window band selection method based on variable starting wavelength and wavelength number in S7 is used for wavelength selection in the first stage; Based on the wavelength selection in the first stage, a multi-wavelength phase-out wavelength selection method is used for the wavelength selection in the second stage. The first and second stages are based on the optimal wavelength combination of the two models to determine the modeling effect.

[0018] Preferably, the wavelength combination used for APC quantitative analysis in step S7 includes 31 wavelengths: 1276nm, 1285nm, 1295nm, 1304nm, 1371nm, 1410nm, 1477nm, 1572nm, 1611nm, 1620nm, 1630nm, 1697nm, 1707nm, 1860nm, 1927nm, 1965nm, 2003nm, 2042nm, 2051nm, 2090nm, 2099nm, 2109nm, 2118nm, 2128nm, 2137nm, 2204nm, 2214nm, 2252nm, 2262nm, 2271nm, and 2338nm.

[0019] Preferably, the wavelength combination used for exceeding the standard in step S7 includes 15 wavelengths: 998nm, 1017nm, 1055nm, 1094nm, 1113nm, 1151nm, 1161nm, 1189nm, 1218nm, 1237nm, 1324nm, 1343nm, 1352nm, 1362nm, and 1381nm.

[0020] Preferably, in step S3, three wavelengths of 1103nm, 1448nm, and 1601nm are selected in the NIR double spectrum region of 1000~1700nm to construct a pseudo-color image and extract the average spectrum within the four regions of interest (ROIs).

[0021] The advantages of this invention compared to the prior art are: This invention proposes a near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices, enabling quantitative analysis and identification of excessive levels of APC in salmon slices. This method requires no sample destruction or reagents and can achieve real-time and rapid detection through non-contact imaging. This invention integrates various steps, including salmon slice sample collection, preparation and microbial colony count detection, spectral imaging measurement, region of interest identification and spectral data extraction, training-prediction-external testing design, quantitative analysis and out-of-specification discrimination primitive algorithms, spectral preprocessing, wavelength selection, and model external testing, to achieve integrated innovation of the method. Its core is the selection of wavelength combinations for quantitative analysis and out-of-specification discrimination of salmon slice APC using NIR-HSI technology. The few-wavelength combination model proposed in this invention can provide a solution for developing a small-scale dedicated NIR-HSI detector for salmon fillet detection; This invention addresses the issue that raw salmon slices, which are sold in large quantities, are easily contaminated and spoiled by microorganisms. It proposes a novel method for real-time detection of microbial colony counts based on near-infrared-hyperspectral imaging, which is of great significance for food safety testing. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the method of the present invention.

[0023] Figure 2 A partial image (50×50 pixels) of the APC distribution of salmon fillet samples exposed to air at nine time points (0h, 3h, 6h, 9h, 12h, 15h, 18h, 21h, 24h). Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings.

[0025] The wavelength selection for the near-infrared hyperspectral imaging method for real-time detection of microbial colony counts in salmon fillets includes the following steps: S1. Salmon slice sample collection, preparation, and microbial colony count detection: Salmon slice samples were collected and exposed to air. Sampling was performed according to the air exposure time, and the APC of the samples was determined using the microbial plate count method. The number of visible colonies formed after culturing single or aggregated live microorganisms on the surface of a standard solid culture medium was defined as colony forming units (CFU). The commonly used logarithm of CFU, lg (CFU), was used as the reference physicochemical value for spectral modeling. S2. Spectral Imaging Measurement: Near-infrared hyperspectral imager (spectral range 883-2492 nm) was used to acquire NIR-HSI data of salmon slices. At the same time point, four samples were placed separately on a white background paper for reference, placed on a moving stage, and the stage was moved at 10 mm / s to acquire NIR-HSI data. S3. Region of Interest (ROI) Identification and Spectral Data Extraction: HSI data at three wavelengths were selected in the NIR double spectrum region (1000-1700 nm) to construct a pseudo-color image of the salmon slice sample, making it highly recognizable against the white background paper; using the Region of Interest (ROI) selection tool of ENVI 5.6 software, combined with visual interpretation, four valid contour ROIs were selected; the average spectrum of all pixels of the ROI was calculated for subsequent modeling and verification; S4. Training-Prediction-External Validation Design: Randomly divide the samples into training set, prediction set, and independent external validation set; the training set and prediction set are used for modeling and parameter optimization, while the independent external validation set, which does not participate in modeling, is used to evaluate the model; S5. Primitive Algorithms for Quantitative Analysis and Exceedance Detection: Partial Least Squares (PLS) is used as the primitive algorithm to establish a quantitative analysis model for APC (Active Polymorphic Crush) in salmon slices. Partial Least Squares-Discriminant Analysis (PLS-DA) is used as the classifier primitive algorithm to establish a spectral identification model for APC exceeding the standard in salmon slices. S6. Spectral preprocessing: Standard normal variable (SNV) transformation and multi-mode Norris derivative filtering (NDF) were used for spectral preprocessing; preprocessing parameters were optimized based on the modeling performance of PLS ​​and PLS-DA models. S7. Wavelength Selection: Based on spectral preprocessing, the moving window (MW) band selection method based on variable starting wavelength and wavelength number is used for the first stage of wavelength selection; based on the first stage of wavelength selection, the multi-wavelength elimination (MWSP) wavelength selection method is used for the second stage of wavelength selection; both stages optimize wavelength combinations based on the modeling effects of PLS ​​and PLS-DA models. S8. External Model Validation: The optimal model is validated using the spectra of salmon fillet samples from an external, independent validation set that is not involved in the modeling process, thus completing the method evaluation.

[0026] This patent uses near-infrared hyperspectral imaging (NIR-HSI) as an example to quantitatively analyze and identify excessive microbial colony counts (APC) in salmon slices. It systematically introduces the implementation methods for various stages, including salmon slice sample collection, preparation, microbial colony count detection, spectral imaging measurement, region of interest identification and spectral data extraction, training-prediction-external testing design, basic algorithms for quantitative analysis and excessive detection, spectral preprocessing, wavelength selection, and model external testing. The patent also integrates and innovates these methods. Partial least squares (PLS) and partial least squares-discriminant analysis (PLS-DA) models are used as basic algorithms for spectral quantitative and discriminant analysis. Standard normal variable (SNV) transformation and Norris derivative filtering (NDF) are used for spectral preprocessing. Moving window (MW) combined with multi-wavelength progressive elimination (MWSP) is used for wavelength selection. The patent details the implementation methods and effects of wavelength selection in the near-infrared hyperspectral imaging method for real-time detection of microbial colony counts in salmon slices. However, the implementation methods of this invention are not limited to these specific examples.

[0027] The specific implementation steps are as follows: S1. Salmon slice sample collection, preparation, and microbial colony count detection. Fresh raw salmon fillets (Norwegian Atlantic salmon, approximately 25 g / fillet) from the same batch were collected through legitimate commercial channels. The salmon fillet samples were exposed to air (24°C) from 0h to 24h, with 4 salmon fillet samples taken every hour, for a total of 100 samples.

[0028] The total bacterial count (APC) of salmon fillet samples was determined using the plate count method. Materials included salmon fillet samples, sterile physiological saline, and agar medium. The procedure included sample testing, dilution, incubation, and plate counting. The common logarithm of colony forming units (CFU), lg(CFU), was used as the reference physicochemical value for spectral calibration. The threshold for APC exceeding the limit for salmon fillets was lg(CFU) = 6. All samples were divided into safe (negative, lg(CFU) < 6, 31 cells) and excessive (positive, lg(CFU) ≥ 6, 68 cells), with a total of 99 samples (1 sample was contaminated during the experiment).

[0029] S2, Spectral Imaging Measurement An indoor short-wave infrared hyperspectral imager (Micro-HyperspecSWIR M384, Headwall, USA) with a reflective convex holographic diffraction grating was used. The spectral scanning range was 883–2492 nm, with a wavelength interval of 9.6 nm and 169 spectral channels. The sensor was a mercury cadmium telluride detector (MCT). Spectra were measured at room temperature (25±1℃) and relative humidity (46±1%). Four samples were simultaneously placed on a sheet of white paper (with a small distance between samples to avoid overlap) and placed on a moving stage. The stage moved at 10 mm / s to acquire hyperspectral imaging HSI data.

[0030] S3, Region of Interest Identification and Spectral Data Extraction In order to extract effective spectral information from the HSI data of salmon slices, the region of interest (ROI) of the effective outline of salmon slices was determined using ENVI 5.6 software: (1) Pseudo-color image generation: Considering that the salmon slice samples have high protein and water content, these components have significant absorption in the NIR double spectrum region (1000–1700 nm). Manual combination selection was performed in this spectral region, and HSI data (reflectivity) of three bands of 1103, 1448 and 1601 nm were used to construct a pseudo-color image of the salmon slice samples, so that it has high recognition relative to the background (white paper). The nonlinear intensity stretching algorithm based on square root of ENVI 5.6 was used to enhance the image contrast. (2) Selection of ROI for effective contours: Using the ROI selection tool in ENVI, combined with visual interpretation, pixels with significant spatial variations in reflectance values ​​were selected as the manually selected boundaries for the salmon slice samples, while avoiding shaded areas, to obtain the effective 1st ROI; to reduce the impact of edge effects and shaded areas on spectral extraction, a smaller fixed step size was used to shrink inward based on the initial boundary, to obtain the 2nd, 3rd and 4th ROIs in sequence, for a total of 4 ROIs for each sample. (3) Spectral data extraction: Since the plate colony counting method can provide the APC per milliliter of salmon slice homogenate, which corresponds to the average APC value of all pixels, the average spectrum of all pixels within the ROI should be used as the effective spectral data corresponding to the APC (converting reflectance to absorbance). Based on 4 ROIs for each sample, a total of 396 average spectra were obtained from 99 samples for subsequent modeling and validation.

[0031] S4. Design and evaluation metrics for training-prediction-external testing A rigorous training-prediction-external validation sample experimental design process is adopted. Samples are divided into training, prediction, and validation sets. The training and prediction phases are used for model building and optimization; in the external validation phase, independent samples not involved in modeling are used to validate the model.

[0032] (1) Quantitative analysis of APC: The mean spectrum of the first ROI region of all samples (99 samples) was used, and 25 samples were randomly selected as the test set; the remaining 74 samples were used as the modeling set. The samples in the modeling set were sorted by lg (CFU) value from smallest to largest and divided into training set (49 samples) and prediction set (25 samples), for a total of 5 times. This ensured that the distribution of lg (CFU) values ​​of APC index of the training set and prediction set had a certain similarity (mean and standard deviation were close), and the model established was objective. For each division i Modeling and calculating the root mean square error of prediction and correlation coefficient , Further calculate the corresponding mean and standard deviation ( ). , , , ), = It is used as a comprehensive evaluation index for multi-partition modeling, and based on the minimum Optimize modeling parameters to ensure model stability.

[0033] In the external verification phase, test samples not involved in modeling are used to verify the selected model, and the root mean square error of prediction (SEP) and correlation coefficient (R) are calculated. P Furthermore, the relative root mean square error R-SEP for the test sample with respect to the average measured value C̄ is calculated, where R-SEP = SEP / C̄.

[0034] (2) APC Exceedance Judgment: Based on the exceedance threshold (lg (CFU)=6), all samples were divided into two categories: exceeding the limit (positive, 68 samples) and safe (negative, 31 samples). The two sample groups, exceeding the limit and safe, were randomly assigned to the training, prediction, and test sets, respectively. The sample categories, number of samples, and number of spectra are shown in Table 1. In order to maintain the balance of data volume between the two types of samples in the training and prediction stages, the negative and positive samples used 4 and 2 ROI average spectra, respectively; in the external independent test stage, both types of samples used 3 ROI average spectra.

[0035] Table 1. Sample categories, number of samples, and number of spectra for training, prediction, and external test sets.

[0036] The evaluation metrics for the out-of-range discrimination model are as follows: In the modeling phase, for the training set, prediction set, and modeling set, the identification accuracy of negative samples, positive samples, and total samples is calculated respectively. , , , , , , , , ; %), the calculation formula is as follows: , , (1) , , (2) , , (3) in, , , and These represent the actual number of negative and positive samples in the training and prediction sets, respectively; , , and These represent the number of negative and positive sample spectra correctly identified in the training and prediction sets, respectively. The identification accuracy is based on the total number of samples modeled (…). The global optimal model is selected by finding the maximum value of ).

[0037] During external testing, the selected model is tested using test samples that were not involved in modeling, and the test identification accuracy of negative samples, positive samples, and total samples is calculated based on the true category of the test samples. , , The calculation formula is as follows: , , (4) It is noted that in the external testing process, data from three spectra were used for both samples that exceeded the standard and those that were safe. The comprehensive judgment results of the three spectra were used to determine the category of the sample by a "best of three" voting method, and the sample identification accuracy was calculated.

[0038] S5, Quantitative Analysis and Exceedance Detection Basic Algorithm Refer to S5 in the "Summary of the Invention".

[0039] S6, Spectral Preprocessing Refer to S6 in the "Summary of the Invention". The parameters of the Norris derivative filter (NDF) are set as follows: derivative order. ;Smooth points Difference interval number . S7, Wavelength Selection Three wavelength selection methods were used to establish quantitative and discriminant models for spectral analysis. Finally, the optimal wavelength selection method was determined based on the modeling results.

[0040] The Moving Window (MW) Band Selection method is based on Moving Window-Partial Least Squares (MW-PLS) and Moving Window-Partial Least Squares-Discriminant Analysis (MW-PLS-DA). The initial wavelength is used ( I ) and number of wavelengths ( N Using the search parameters, all sub-bands within the wavelength range are traversed for modeling. A PLS-based quantitative spectral model and a PLS-DA-based spectral discrimination model are established respectively. The quantitative performance of the prediction set is then evaluated using the SEP (Search Equation for Predictive Spectroscopy). + And the discriminant effect RAR PThe optimal wavelength band is determined. In the current embodiment, the wavelength search range is 883-2492 nm, the total number of wavelengths is 169, and the parameters are set as follows: , .

[0041] Multi-wavelength phase-out (MWSP) based methods include Multi-wavelength phase-out with partial least squares (MWSP-PLS) and Multi-wavelength phase-out with partial least squares discriminant analysis (MWSP-PLS-DA). Within the existing wavelength range, phase-out is performed in ascending order of wavelength size. m A continuous wavelength ( The remaining wavelength combinations were used for modeling, and a PLS-based quantitative spectral model and a PLS-DA-based spectral discrimination model were established respectively. The quantitative performance of the prediction set was then evaluated using the SEP (Sequential Estimation and Precision). + And the discriminant effect RAR P To determine the modeling effect that yields the best results for the remaining wavelength combinations. m One wavelength was eliminated; based on this, the same method was used to eliminate wavelengths again. m A series of consecutive wavelengths, and so on, are gradually eliminated until only a few remain. n wavelengths ( n ≤ m Then, from the best model in each round of elimination, the corresponding model is selected. m The optimal model; then from all corresponding m The globally optimal model is determined from the optimal model. In the current embodiment, the parameters are set as follows: , M =10.

[0042] The MW band selection method with forward optimization was used for wavelength selection in the first stage; the MWSP wavelength selection method with backward optimization was used for wavelength selection in the second stage; and the moving window-multi-wavelength stepwise elimination-partial least squares (MW-MWSP-PLS) method and the moving window-multi-wavelength stepwise elimination-partial least squares-discriminant analysis (MW-MWSP-PLS-DA) method with forward and backward optimization functions were established.

[0043] S8. APC quantitative analysis modeling for salmon fillet samples Based on the spectral preprocessing and wavelength selection methods mentioned in S6 and S7, there are two spectral preprocessing options: SNV and NDF; and three wavelength selection options: MW, MWSP, and MW-MWSP. Correspondingly, there are six combined models of spectral preprocessing and wavelength selection. MW-PLS, MWSP-PLS, and MW-MWSP-PLS models for APC quantification of salmon slices were established (see Table 2). The MW-MWSP-PLS model based on SNV preprocessed spectra showed the best prediction performance, with the optimal wavelength combination being: 1276, 1285, 1295, 1304, 1371, 1410, 1477, 1572, 1611, 1620, 1630, 1697, 1707, 1860, 1927, 1965, 2003, 2042, 2051, 2090, 2099. 2109, 2118, 2128, 2137, 2204, 2214, 2252,2262, 2271, 2338 (nm, N =31).

[0044] Table 2. Modeling parameters and prediction results of APC quantitative MW-PLS, MWSP-PLS, and MW-MWSP-PLS models.

[0045] Modeling for APC Exceedance Detection in S9 Salmon Slice Samples Based on the spectral preprocessing and wavelength selection methods mentioned in S6 and S7, there are two spectral preprocessing options: SNV and NDF; and three wavelength selection options: MW, MWSP, and MW-MWSP. Correspondingly, there are six combined models for spectral preprocessing and wavelength selection. MW-PLS-DA, MWSP-PLS-DA, and MW-MWSP-PLS-DA models (see Table 3) were established for judging APC exceedance in salmon slice samples. The MW-MWSP-PLS-DA model based on SNV preprocessed spectra showed the best prediction performance, and the corresponding optimal wavelength combination is: 998, 1017, 1055, 1094, 1113, 1151, 1161, 1189, 1218, 1237, 1324, 1343, 1352, 1362, 1381 (nm). N =15).

[0046] Table 3 Modeling parameters and discrimination results of MW-PLS-DA, MWSP-PLS-DA, and MW-MWSP-PLS-DA models for APC exceedance discrimination.

[0047] Quantitative analysis of APC: The MW-MWSP-PLS model based on SNV preprocessed spectra was tested using the spectra of salmon fillet samples from an external independent test set that were not involved in the modeling (see Table 4). The correlation coefficient R between the spectral predictions and measured values ​​of salmon fillet APC is shown. P With a mean squared error (R-SEP) of 0.891 and a relative error (R-SEP) of 9.5%, the NIR-HSI method demonstrates good accuracy in real-time detection of total microbial colonies in salmon fillet samples.

[0048] Visualization of APC distribution at sites on the surface of salmon fillet samples: Using the optimal quantitative analysis model and the spectrum of each pixel in the NIR-HSI image of salmon fillet samples, the total number of microbial colonies at sites on the surface of salmon fillet samples can be predicted, and an APC distribution map can be plotted. This is done according to air exposure time (0h, 3h, 6h, 9h, 12h, 15h, 18h, 21h, 24h). Figure 2 The local spatial distribution of total microbial colonies (50×50 pixels) of nine salmon slice samples is shown. Pixels with lg (CFU) < 6 are represented in green, pixels with 6 ≤ lg (CFU) ≤ 8 are represented in blue, and pixels with lg (CFU) > 8 are represented in red. It can be observed that the degree of microbial contamination of the salmon slice samples increases with the increase of air exposure time.

[0049] Table 4. External independent test results of the optimal MW-MWSP-PLS model

[0050] APC Exceedance Detection: The MW-MWSP-PLS model based on SNV preprocessed spectra was tested using the spectra of salmon fillet samples from an external independent test set that were not involved in the modeling (see Table 5). The overall identification accuracy (RAR) for salmon fillet APC detection was... V With a success rate of 96.0%, the NIR-HSI method demonstrates high accuracy in real-time detection of excessive total microbial colonies in salmon fillet samples.

[0051] Table 5. External independent test results of the optimal MW-MWSP-PLS-DA model

[0052] Experimental results show that near-infrared hyperspectral imaging can quantitatively analyze and discriminate between total microbial colonies and exceedances in salmon fillets, and can visualize the APC distribution at sites on the surface of salmon fillet samples. The proposed few-wavelength combination model can provide a solution for developing a small-scale dedicated NIR-HSI detector for salmon fillet testing. This method does not require sample destruction or reagents, and can achieve real-time rapid detection through non-contact imaging. The detection and modeling method proposed in this embodiment is also expected to be applied to the detection of total microbial colonies in other meat samples.

[0053] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. This invention discloses a near-infrared hyperspectral imaging method for real-time detection of microbial colony counts in salmon fillets, comprising the following steps: S1: Prepare salmon samples that have been exposed to air for different periods of time and test for microbial colony count; S2: Perform spectral imaging measurements and acquire hyperspectral imaging HSI data; S3: Based on the data collected in S2, identify the region of interest and extract spectral data; S4: Divide the samples into a training set, a prediction set, and an independent external test set; S5: Establish a quantitative spectral analysis model for APC in salmon fillets and a spectral identification model for judging APC exceeding the standard in salmon fillets; S6: Perform standard normal variable transformation (SNV) and Norris derivative filtering (NDF) preprocessing on the spectral data; S7: Optimize wavelength combinations through a two-stage wavelength selection method involving moving windows and multi-wavelength phase-out; S8: Verify model accuracy using an independent test set. The advantages of this invention compared to existing technologies are: it provides a near-infrared hyperspectral imaging method for real-time detection of microbial colonies in salmon fillets, offering a solution for developing a small-scale dedicated NIR-HSI detector for salmon fillet detection using a few-wavelength combination model.

2. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 1, characterized in that: S1 includes exposing salmon slices to air for 0-24 hours, taking 4 slices per hour, and determining the total bacterial count (APC) using the microbial plate count method.

3. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 1, characterized in that: In S2, a near-infrared hyperspectral imager with a spectral range of 883~2492nm is used; four samples are placed separately on a white background paper for reference, and then placed on a moving stage. The stage moves at 10mm / s to collect HSI data.

4. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 2, characterized in that: S3 includes constructing a pseudo-color image in the NIR double spectral region to identify the salmon slice region and extracting the average spectrum within multiple regions of interest (ROIs).

5. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 1, characterized in that: In S5, partial least squares is used as the primitive algorithm to establish a quantitative analysis model of salmon slice APC spectrum; partial least squares-discriminant analysis is used as the classifier primitive algorithm.

6. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 5, characterized in that: S6 includes preprocessing using Standard Normal Variable Transform (SNV) and Norris Derivative Filter (NDF), and selecting preprocessing parameters based on the modeling effects of the two models.

7. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 5, characterized in that: The moving window band selection method based on variable starting wavelength and wavelength number in S7 is used for wavelength selection in the first stage. Based on the wavelength selection in the first stage, a multi-wavelength phase-out wavelength selection method is used for the wavelength selection in the second stage. The first and second stages are based on the optimal wavelength combination of the two models to determine the modeling effect.

8. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 7, characterized in that: The wavelength combination used for APC quantitative analysis in step S7 includes 31 wavelengths: 1276nm, 1285nm, 1295nm, 1304nm, 1371nm, 1410nm, 1477nm, 1572nm, 1611nm, 1620nm, 1630nm, 1697nm, 1707nm, 1860nm, 1927nm, 1965nm, 2003nm, 2042nm, 2051nm, 2090nm, 2099nm, 2109nm, 2118nm, 2128nm, 2137nm, 2204nm, 2214nm, 2252nm, 2262nm, 2271nm, and 2338nm.

9. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 7, characterized in that: The wavelength combination used for exceeding the standard in step S7 includes 15 wavelengths: 998nm, 1017nm, 1055nm, 1094nm, 1113nm, 1151nm, 1161nm, 1189nm, 1218nm, 1237nm, 1324nm, 1343nm, 1352nm, 1362nm, and 1381nm.

10. The near-infrared hyperspectral imaging method for real-time detection of microbial colony count in salmon slices according to claim 4, characterized in that: In step S3, three wavelengths of 1103nm, 1448nm, and 1601nm are selected in the NIR double spectrum region of 1000~1700nm to construct a pseudo-color image, and the average spectrum within the four regions of interest (ROIs) is extracted.

Images