A method and system for evaluating the freshness of fish meat

Abstract

The present application relates to the technical field of food quality detection, in particular to a fish freshness evaluation method and system, comprising: obtaining hyperspectral reflectance curve, surface texture microscopic image and volatile odor substance concentration spectrum of the fish to be measured, extracting protein degradation characteristic spectral absorption band, microstructure integrity multi-level entropy set and microbial metabolism indicative odor fingerprint. The three types of characteristics are input into an improved fusion evaluation algorithm, which adopts a weighted decision fusion framework, introduces a time decay factor dynamic weight distribution and a feature reliability adaptive correction mechanism, and outputs a dynamic freshness quantitative index and a shelf life stage label through feature-level fusion and time series evolution analysis. The present application can match the time sequence change of fish quality, weaken the interference of characteristic fluctuation, and improve the stability and accuracy of the evaluation result.

Description

Technical Field

[0001] This invention relates to the field of food quality testing technology, and in particular to a method and system for evaluating the freshness of fish. Background Technology

[0002] Current methods for detecting the freshness of fish meat mostly employ multi-source data acquisition methods, including hyperspectral analysis, microscopic texture analysis, and volatile odor analysis, and use fixed-weight feature fusion algorithms to determine freshness. Conventional solutions perform static weighted calculations on spectral features, texture features, and odor features, relying on constant coefficients to complete the decision output, which can only achieve static index detection and grading.

[0003] Static weighted fusion cannot adapt to the time-series changes in the fish spoilage process. Different features exhibit varying degrees of reliability decay over storage time, and fixed weights cannot match the real-time effectiveness of each feature. Single-class features are susceptible to data fluctuations caused by the detection environment and sample differences, and there is a lack of correction methods for feature reliability. The temporal coupling relationship between spectral, texture, and odor features is not utilized, resulting in discrepancies between the fusion results and the actual evolution of fish freshness, and insufficient accuracy in shelf-life segmentation.

[0004] The current state of constant fusion weights needs to be changed, and a dynamic weight allocation model based on time decay factors needs to be constructed. Fusion bias caused by fluctuations in feature credibility needs to be eliminated, and an adaptive feature credibility correction process needs to be established, combining feature-level fusion with time-series evolution analysis to achieve dynamic evaluation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and to propose a method and system for evaluating the freshness of fish.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for evaluating the freshness of fish meat, comprising:

[0007] Acquire initial multi-source sensor data of the fish meat to be tested, including hyperspectral reflectance curve, surface texture microscopic image and volatile odor substance concentration spectrum;

[0008] Wavelength-reflectance correlation analysis was performed on the hyperspectral reflectance curve to extract the characteristic spectral absorption bands that characterize the degradation state of fish myofibril protein.

[0009] Multi-scale texture decomposition was performed on the surface texture microscopic image to calculate a multi-level entropy value set characterizing the integrity of the microstructure of fish meat tissue.

[0010] Principal component dimensionality reduction and pattern matching were performed on the concentration spectrum of the volatile odor substances to identify indicative odor fingerprints associated with microbial metabolic activities;

[0011] The characteristic spectral absorption bands, the multi-level entropy value set, and the indicative odor fingerprint are input into the improved fish freshness fusion evaluation algorithm;

[0012] The improved fish freshness fusion evaluation algorithm is used to perform feature-level fusion and time series evolution analysis on the characteristic spectral absorption band, the multi-level entropy value set and the indicative odor fingerprint, and outputs the dynamic freshness quantification index and shelf life stage label of the fish to be tested.

[0013] The improved fish freshness fusion evaluation algorithm adopts a weighted decision fusion architecture. The improvement is achieved by introducing a dynamic weight allocation mechanism based on time decay factor and an adaptive correction process based on feature credibility.

[0014] As a further aspect of the present invention, wavelength-reflectance correlation analysis is performed on the hyperspectral reflectance curve to extract characteristic spectral absorption bands characterizing the degradation state of fish myofibril proteins, including:

[0015] The hyperspectral reflectance curve is smoothed and denoised to eliminate abnormal fluctuations in reflectance caused by equipment noise and ambient stray light.

[0016] On the smoothed and denoised hyperspectral reflectance curve, the first derivative spectrum of each wavelength point is calculated, and the wavelength position where the reflectance change rate reaches a local maximum is identified and marked as a candidate absorption peak.

[0017] Based on the fish biochemical component database, candidate absorption peaks located in the characteristic absorption bands of myofibril proteins are screened from the candidate absorption peaks to form a preliminary set of spectral absorption bands.

[0018] Calculate the absorption intensity, full width at half maximum (FWHM), and waveform symmetry of each absorption band in the preliminary spectral absorption band set;

[0019] The absorption intensity, full width at half maximum (FWHM), and waveform symmetry are compared with the pre-stored standard fresh fish spectral absorption band characteristics. The absorption bands with similarity exceeding the threshold are identified as characteristic spectral absorption bands representing the degradation state of myofibril proteins.

[0020] As a further aspect of the present invention, the surface texture microscopic image is subjected to multi-scale texture decomposition to calculate a multi-level entropy value set characterizing the integrity of the microstructure of fish flesh tissue, including:

[0021] The surface texture microscopic image is converted to grayscale and the region of interest containing fish muscle fiber bundles and connective tissue is segmented.

[0022] Wavelet transform is used to decompose the image of the region of interest into multiple scales to obtain sub-band images containing horizontal, vertical and diagonal directional information at different scales;

[0023] Calculate the energy entropy of each sub-band image at each scale, whereby the energy entropy reflects the complexity and irregularity of the image texture at each scale;

[0024] The energy entropy calculated at all scales is arranged and combined in order from coarse to fine scale to generate the multi-level entropy value set.

[0025] The gradient of entropy values ​​as a function of scale is extracted from the multi-level entropy value set. This gradient is used to quantify the degree of degradation of the microstructure of fish meat at different scales.

[0026] As a further aspect of the present invention, principal component dimension reduction and pattern matching are performed on the concentration spectrum of the volatile odor substances to identify indicative odor fingerprints associated with microbial metabolic activities, including:

[0027] The concentration spectrum of the volatile odor substances is standardized and preprocessed to eliminate the influence of differences in the concentration dimensions between different odor substances;

[0028] Principal component analysis was applied to the concentration spectrum of volatile odor substances after standardization pretreatment to calculate the eigenvalues ​​and eigenvectors of its covariance matrix;

[0029] Based on the preset cumulative variance contribution rate threshold, the number of principal components is selected, and the original high-dimensional odor concentration data is projected onto the feature subspace composed of the selected principal components to obtain the dimensionality-reduced principal component score data.

[0030] Using a pre-established fish meat spoilage stage odor database, which contains clustering center patterns of samples at different spoilage stages in the principal component feature subspace;

[0031] Calculate the similarity distance between the dimensionality-reduced principal component score data and the cluster center patterns of each spoilage stage in the fish spoilage stage odor database;

[0032] The feature vectors corresponding to the putrefaction stage cluster centers with the smallest similarity distance to the reduced principal component score data are combined to determine the indicative odor fingerprint that characterizes the current sample's microbial metabolic activity.

[0033] As a further aspect of the present invention, the step of performing feature-level fusion and time-series evolution analysis on the characteristic spectral absorption bands, the multi-level entropy value set, and the indicative odor fingerprint using the improved fish freshness fusion evaluation algorithm includes:

[0034] The characteristic spectral absorption bands are mapped into a continuous numerical sequence characterizing the degree of protein degradation;

[0035] The multi-level entropy value set is mapped into a continuous numerical sequence characterizing the degree of tissue structure deterioration;

[0036] The indicative odor fingerprint is mapped into a continuous numerical sequence characterizing the microbial putrefaction process;

[0037] The continuous numerical sequences characterizing the degree of protein degradation, the continuous numerical sequences characterizing the degree of tissue structural deterioration, and the continuous numerical sequences characterizing the microbial putrefaction process are each subjected to independent timestamp alignment and interpolation processing to form a time-synchronized fused feature time series.

[0038] A sliding window analysis is performed on the fusion feature time series. Within each window, a dynamic freshness metric is calculated based on the weighted decision fusion architecture of the improved fish freshness fusion evaluation algorithm. The value of the dynamic freshness metric at the center of the window is then output.

[0039] As a further aspect of the present invention, the dynamic weight allocation mechanism based on the time decay factor in the improved fish freshness fusion evaluation algorithm includes:

[0040] An initial weight is defined for the continuous numerical sequence characterizing the degree of protein degradation, the continuous numerical sequence characterizing the degree of tissue structural deterioration, and the continuous numerical sequence characterizing the microbial putrefaction process, respectively.

[0041] Based on the biochemical dynamics model of fish spoilage, the time point and rate of change of the indicator represented by each continuous numerical sequence are preset during the spoilage process.

[0042] During the sliding window analysis, a time decay factor is calculated based on the time difference between the current analysis time and the starting time of each indicator. The time decay factor increases as the time difference increases.

[0043] The time decay factor is multiplied by the initial weight of each sequence to obtain the dynamic weight of each sequence in the fusion calculation at the current time.

[0044] As a further aspect of the present invention, the adaptive correction process based on feature reliability in the improved fish freshness fusion evaluation algorithm includes:

[0045] For the characteristic spectral absorption band, calculate its matching degree with the pre-stored standard database, and normalize the matching degree to the characteristic confidence degree of the characteristic spectral absorption band;

[0046] For the multi-level entropy value set, the range of entropy value change is calculated based on historical data, and the deviation score is calculated according to the position of the current entropy value within the range of change. The deviation score is then normalized to the feature credibility of the multi-level entropy value set.

[0047] For the indicative odor fingerprint, its principal component score is compared with the Euclidean distance of the center point of the typical putrefaction stage, and the reciprocal of the Euclidean distance is normalized to the feature confidence of the indicative odor fingerprint.

[0048] Based on the feature confidence of the characteristic spectral absorption band, the feature confidence of the multi-level entropy value set, and the feature confidence of the indicative odor fingerprint, the corresponding fusion weight coefficients in the weighted decision fusion architecture are dynamically adjusted.

[0049] Using the adjusted fusion weight coefficients, a weighted fusion calculation is performed on the input characteristic spectral absorption bands, the multi-level entropy value set, and the indicative odor fingerprint.

[0050] As a further aspect of the present invention, the step of calculating the matching degree between the characteristic spectral absorption band and a pre-stored standard database includes:

[0051] From the pre-stored standard database, retrieve the average spectral absorption band and its standard deviation band of standard fish meat samples of different freshness grades under the same characteristic absorption band;

[0052] The waveform similarity of the characteristic spectral absorption band to be tested with the standard spectral absorption band of each freshness grade is measured. The waveform similarity measurement includes correlation coefficient calculation and root mean square error calculation.

[0053] Based on the correlation coefficient and the root mean square error, and combined with the standard deviation band corresponding to the freshness grade, calculate the confidence probability that the absorption band of the spectral feature to be tested falls into the standard distribution of the freshness grade.

[0054] The confidence probability value corresponding to the freshness level with the highest confidence probability is selected as the matching degree between the characteristic spectral absorption band and the pre-stored standard database.

[0055] As a further aspect of the present invention, the step of outputting the dynamic freshness quantification index and shelf-life stage label of the fish meat to be tested includes:

[0056] The dynamic freshness metric is a continuous value between a preset minimum value and a preset maximum value, where the preset minimum value corresponds to a completely rotten state and the preset maximum value corresponds to an absolutely fresh state.

[0057] Multiple threshold ranges are preset, and each threshold range corresponds to a shelf life stage label. The shelf life stage label includes the fresh stage, the quality decline stage, the early spoilage stage, and the inedible stage.

[0058] The calculated dynamic freshness metric is compared with a preset threshold range to determine the threshold range to which it belongs.

[0059] Assign the shelf life stage label corresponding to the threshold interval to the fish meat to be tested;

[0060] The output includes a freshness evaluation report containing the numerical values ​​of the dynamic freshness metric and the label name of the shelf life stage.

[0061] As a further aspect of the present invention, the present invention also includes a fish freshness evaluation system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the fish freshness evaluation method described above.

[0062] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0063] The weighted decision fusion architecture incorporates a dynamic weight allocation mechanism based on a time decay factor, adjusting the fusion weights of characteristic spectral absorption bands, multi-level entropy sets, and indicative odor fingerprints according to the fish storage time series. This aligns with the temporal evolution of myofibril protein degradation, microstructural damage, and odor substance release, matching the actual contribution of each feature at different storage stages. It weakens the weights of features whose effectiveness decreases with spoilage, strengthens the proportion of features with increased freshness correlation, and ensures that the weight distribution remains synchronized with the changes in fish quality.

[0064] The feature-based adaptive correction process can adjust the reliability parameters of three types of input features in real time, offsetting environmental interference and systematic errors during hyperspectral acquisition, texture microscopy imaging, and odor substance detection. It corrects the discrete bias of feature data, ensuring the stability and effectiveness of the feature data involved in the fusion, and reducing the interference of single-type feature anomalies on the overall evaluation results. Feature-level fusion and time-series evolution analysis are mutually compatible, and the output dynamic freshness quantification index can continuously reflect the changes in fish meat quality, with shelf-life stage labeling closely matching actual storage stage changes. Attached Figure Description

[0065] Figure 1 This is a flowchart of a method for evaluating the freshness of fish meat according to the present invention;

[0066] Figure 2 A flowchart of a method for extracting characteristic spectral absorption bands;

[0067] Figure 3 A flowchart illustrating the method for calculating a multi-level entropy set. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0069] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0070] See Figure 1 This invention provides a method for evaluating the freshness of fish meat. The method uses the hyperspectral reflectance curve, surface texture microscopic image, and volatile odor compound concentration spectrum of the fish meat as initial multi-source sensing data. The hyperspectral reflectance curve is analyzed to extract characteristic spectral absorption bands representing the degradation state of myofibril proteins in the fish meat. The surface texture microscopic image is analyzed to calculate a multi-level entropy set representing the integrity of the microstructure of the fish meat tissue. The volatile odor compound concentration spectrum is analyzed to identify indicative odor fingerprints associated with microbial metabolic activity. These characteristic spectral absorption bands, multi-level entropy sets, and indicative odor fingerprints are input into an improved fish meat freshness fusion evaluation algorithm. This improved fish meat freshness fusion evaluation algorithm adopts a weighted decision fusion architecture, and its improvement lies in introducing a dynamic weight allocation mechanism based on a time decay factor and an adaptive correction process based on feature credibility. The algorithm performs feature-level fusion and time-series evolution analysis on the input features, and finally outputs a dynamic freshness quantification index and shelf-life stage label for the fish meat being tested.

[0071] In one embodiment of the present invention, see [reference] Figure 2Wavelength-reflectance correlation analysis was performed on the hyperspectral reflectance curve to extract characteristic spectral absorption bands representing the degradation state of fish myofibrillar proteins. This process included smoothing and denoising the hyperspectral reflectance curve to eliminate abnormal fluctuations in reflectance caused by equipment noise and stray light from the environment. On the smoothed and denoised hyperspectral reflectance curve, the first derivative spectrum at each wavelength point was calculated to identify the wavelength positions where the rate of change of reflectance reached a local maximum; these positions were marked as candidate absorption peaks. Based on an existing database of fish biochemical components, candidate absorption peaks located in the characteristic absorption bands of myofibrillar proteins were selected from the candidate absorption peaks, forming a preliminary set of spectral absorption bands. The absorption intensity, full width at half maximum (FWHM), and waveform symmetry of each absorption band in the preliminary set were calculated. The calculated absorption intensity, FWHM, and waveform symmetry were compared item by item with the pre-stored characteristics of standard fresh fish spectral absorption bands. Absorption bands with similarity exceeding a preset threshold were determined as the final characteristic spectral absorption bands used to characterize the degradation state of myofibrillar proteins.

[0072] In practice, wavelength-reflectance correlation analysis is performed on the hyperspectral reflectance curve to extract characteristic spectral absorption bands representing the degradation state of fish myofibrillar proteins. This process begins with smoothing and denoising the hyperspectral reflectance curve. The smoothing and denoising process uses a digital filtering algorithm to eliminate abnormal fluctuations in reflectance caused by equipment noise and ambient stray light. On the smoothed and denoised hyperspectral reflectance curve, the first derivative spectrum is calculated for each wavelength point. The first derivative spectrum is used to identify the wavelength positions where the rate of change of reflectance reaches a local maximum. These wavelength positions are marked as candidate absorption peaks by detecting the sign change and amplitude threshold of the first derivative spectrum. Based on the fish biochemical component database, candidate absorption peaks located in the characteristic absorption bands of myofibrillar proteins are screened from the candidate absorption peaks. These candidate absorption peaks in the characteristic absorption bands of myofibrillar proteins constitute a preliminary set of spectral absorption bands.

[0073] The absorption intensity, full width at half maximum (FWHM), and waveform symmetry of each absorption band in the preliminary spectral absorption band set are calculated. Absorption intensity is defined as the difference between the reflectance at the peak and the baseline reflectance. FWHM is defined as the full width at half the height of the absorption peak. Waveform symmetry is quantified by comparing the area ratio or slope ratio of the left and right halves of the absorption peak. The absorption intensity, FWHM, and waveform symmetry are compared item by item with the pre-stored standard fresh fish spectral absorption band characteristics. This comparison involves calculating the similarity score between the absorption band characteristics to be tested and the standard characteristics. Absorption bands with similarity scores exceeding a preset threshold are identified as characteristic spectral absorption bands representing the degradation state of myofibril proteins.

[0074] In some embodiments, smoothing and denoising can be achieved using moving average filtering or Savitzky-Golay convolution filtering. The parameters of moving average filtering or Savitzky-Golay convolution filtering are set according to the spectral resolution and noise characteristics of the hyperspectral imaging system. The calculation of the first-order derivative spectrum uses the central difference method. The central difference method approximates the derivative at each wavelength point using the reflectance values ​​of adjacent wavelength points. The wavelength position where the reflectance change rate reaches a local maximum must satisfy the condition that the first derivative changes from positive to negative and the change amplitude exceeds the noise floor. Based on the fish biochemical component database, the characteristic absorption bands of myofibril proteins cover specific wavelength ranges. Peaks located within these wavelength ranges are screened from candidate absorption peaks, and spurious peaks that significantly deviate from the known protein absorption characteristics are excluded during the screening process. Each absorption band in the preliminary spectral absorption band set needs to have its absorption intensity, full width at half maximum (FWHM), and waveform symmetry calculated independently. Before calculating the absorption intensity, the accurate vertex and baseline reflectance of the absorption peak need to be determined through interpolation or fitting. The FWHM calculation involves finding the left and right boundary wavelengths at half the height of the absorption peak. The waveform symmetry calculation can use the ratio of the area of ​​the left half to the area of ​​the right half. Optionally, the similarity score in item-by-item comparison can be quantified using a mathematical formula, expressed as follows:

[0075] ;

[0076] in: Represents the similarity score. This indicates the absorption intensity of the absorption band being measured. This represents the standard absorption intensity in the pre-stored standard fresh fish meat spectral absorption band characteristics. This indicates the range of variation in absorption intensity within the training dataset; This represents the full width at half maximum (FWHM) of the absorption band being measured. This represents the standard full width at half maximum (FWHM) in the pre-stored standard fresh fish meat spectral absorption band characteristics. This indicates the range of variation of the full width and half-height in the training dataset. This indicates the waveform symmetry of the absorption band under test. This indicates the standard waveform symmetry in the pre-stored standard fresh fish meat spectral absorption band characteristics. This indicates the range of waveform symmetry variation within the training dataset. Similarity score. The closer the value is to 1, the higher the match between the absorption band being tested and the standard feature. In practice, the preset threshold is determined by analyzing the similarity score distribution of fresh and spoiled samples in historical data. The preset threshold is set as a critical value that can distinguish different degradation states. When the similarity score... When the value is greater than or equal to a preset threshold, the absorption band to be tested is considered to be sufficiently similar to the standard feature, and thus identified as a characteristic spectral absorption band.

[0077] It is understandable that the extraction of characteristic spectral absorption bands depends on the quality of the hyperspectral reflectance curve and the effectiveness of preprocessing. Smoothing and denoising reduce noise interference to improve the accuracy of subsequent derivative analysis and peak identification. The application of first-order derivative spectroscopy can highlight the changing trend of the reflectance curve, which helps to accurately locate the absorption peak position. Screening based on the fish biochemical component database ensures that candidate absorption peaks are related to the target biochemical component myofibrillar protein, avoiding interference from non-specific absorption bands. Absorption intensity, full width at half maximum (FWHM), and waveform symmetry serve as quantitative descriptors for absorption bands, comprehensively reflecting the morphological and intensity characteristics of the absorption bands. Item-by-item comparison with the pre-stored standard fresh fish spectral absorption band features converts the quantitative descriptors into comparable similarity scores, which are objectively evaluated through formula calculation. The introduction of preset thresholds provides a clear standard for judging characteristic spectral absorption bands, ensuring the consistency and repeatability of the extraction process. It is understood that the entire implementation method forms a coherent process from data preprocessing to feature determination, and the final output characteristic spectral absorption bands are used for subsequent fusion evaluation algorithms.

[0078] In one embodiment of the present invention, see [reference] Figure 3 Multi-scale texture decomposition is performed on surface texture microscopic images to calculate a multi-level entropy value set characterizing the integrity of the microstructure of fish flesh tissue. This process includes grayscale conversion of the surface texture microscopic images and segmentation of the region of interest (ROI) containing fish muscle fiber bundles and connective tissue from the processed images. Wavelet transform is used to perform multi-scale decomposition on the ROI image, resulting in sub-band images containing horizontal, vertical, and diagonal directional information at different scales. The energy entropy of each sub-band image at each scale is calculated; this energy entropy reflects the complexity and irregularity of the image texture at the corresponding scale. The energy entropies calculated at all scales are arranged and combined in order of coarse to fine scale to generate the multi-level entropy value set. The gradient of entropy values ​​with scale variation can be further extracted from the generated multi-level entropy value set; this gradient is used to quantify the degree of degradation of the fish flesh tissue microstructure at different scales.

[0079] Principal component analysis (PCA) and pattern matching were performed on the concentration spectrum of volatile odor substances to identify indicative odor fingerprints associated with microbial metabolic activities. This process included standardizing the concentration spectrum of volatile odor substances to eliminate the influence of differences in concentration dimensions among different odor substances. PCA was then applied to the standardized volatile odor concentration spectrum to calculate the eigenvalues ​​and eigenvectors of its covariance matrix. Based on a preset cumulative variance contribution rate threshold, an appropriate number of principal components were selected, and the original high-dimensional odor concentration data was projected onto the feature subspace formed by the selected principal components to obtain the dimensionality-reduced principal component scores. A pre-established odor database of fish spoilage stages, containing cluster center patterns of samples at different spoilage stages in the principal component feature subspace, was used. The similarity distance between the dimensionality-reduced principal component scores and the cluster center patterns of each spoilage stage in the fish spoilage stage odor database was calculated. The eigenvectors corresponding to the spoilage stage cluster centers with the smallest similarity distance to the dimensionality-reduced principal component scores were combined to determine the indicative odor fingerprint characterizing the microbial metabolic activities of the current sample.

[0080] In practice, multi-scale texture decomposition is performed on the surface texture microscopic image to calculate a multi-level entropy set characterizing the integrity of the microstructure of fish flesh tissue. This process begins with grayscale conversion of the surface texture microscopic image, which transforms the color image into a single-channel grayscale image. Regions of interest (ROIs) containing fish muscle fiber bundles and connective tissue are then segmented from the grayscale image. Wavelet transform is used to perform multi-scale decomposition on the ROI image. The wavelet transform uses specific wavelet basis functions to convolve and downsample the image, resulting in sub-band images containing horizontal, vertical, and diagonal information at different scales. The energy entropy of each sub-band image at each scale is calculated. The energy entropy calculation requires first obtaining the sum of squares of all pixel values ​​in the sub-band image as energy. Energy entropy reflects the complexity and irregularity of the image texture at the corresponding scale.

[0081] The energy entropy calculated at all scales is arranged and combined in order from coarse to fine scale, forming an ordered numerical sequence. This ordered numerical sequence generates the multi-level entropy value set. The gradient of entropy values ​​with scale is extracted from the multi-level entropy value set. The gradient of entropy values ​​with scale is obtained by calculating the difference of entropy values ​​between adjacent scales. This gradient is used to quantify the degree of degradation of the microstructure of fish meat at different scales.

[0082] In some embodiments, grayscale processing can be achieved by using a weighted average method to convert the red, green, and blue channel values ​​into grayscale values. Segmentation of the region of interest can be achieved through manual selection or an automatic segmentation algorithm based on thresholding and edge detection. Multi-scale decomposition of the wavelet transform can be performed using a discrete two-dimensional wavelet transform, which generates a low-frequency approximate sub-band and high-frequency detail sub-bands in the horizontal, vertical, and diagonal directions at each decomposition level. The formula for calculating energy entropy can be expressed as...

[0083] ;

[0084] in: Represents energy entropy. This represents the proportion of each pixel's energy value to the total energy of the subband image after normalization. Generating a multi-level entropy set requires pre-setting the number of decomposition levels; the number of levels determines the number of elements in the set. The entropy gradient can be calculated as follows: ,in This represents the entropy gradient from the k-th scale to the (k+1)-th scale. and These represent the energy entropy values ​​at the k-th and (k+1)-th scales in the multi-level entropy value set, respectively.

[0085] Optionally, principal component analysis (PCA) and pattern matching are performed on the volatile odor concentration spectrum to identify indicative odor fingerprints associated with microbial metabolic activities. This process includes standardizing the volatile odor concentration spectrum, which eliminates the influence of differences in concentration dimensions between different odor substances by subtracting the mean and then dividing by the standard deviation. PCA is then applied to the standardized volatile odor concentration spectrum to calculate the eigenvalues ​​and eigenvectors of the covariance matrix. The number of principal components is selected based on a preset cumulative variance contribution rate threshold, typically set to 85% to 95%. The original high-dimensional odor concentration data is projected onto the feature subspace composed of the selected principal components. Projection is achieved by multiplying the original data matrix with the selected principal component eigenvector matrix, yielding the dimensionality-reduced principal component scores. A pre-established fish spoilage stage odor database is used, which contains cluster center patterns of samples at different spoilage stages in the principal component feature subspace. The similarity distance between the dimensionality-reduced principal component score data and the cluster center patterns of each spoilage stage in the fish spoilage stage odor database is calculated. The feature vectors corresponding to the spoilage stage cluster centers with the smallest similarity distance to the dimensionality-reduced principal component score data are combined to determine the indicative odor fingerprint characterizing the microbial metabolic activity of the current sample.

[0086] It is understandable that multi-scale texture decomposition analyzes the texture features of images at different resolutions through wavelet transform, energy entropy quantifies the randomness and complexity of textures at each scale, multi-level entropy sets orderly record the rules of texture feature changes with the observation scale, and entropy gradients further reveal the speed and pattern of texture structure changes. Preprocessing of surface texture microscopic images is fundamental, and accurate region-of-interest segmentation ensures that the analysis focuses on effective fish tissue. Similarly, odor fingerprinting relies on principal component analysis for effective compression of high-dimensional data. Standardization preprocessing ensures equal contributions of variables with different dimensions in dimensionality reduction, the cumulative variance contribution rate threshold balances information preservation and data simplification, and pattern matching with a pre-built database correlates the abstract scores after dimensionality reduction to specific microbial metabolic activity stages. Indicative odor fingerprints are the result of this series of mathematical transformations and comparisons.

[0087] In one embodiment of the present invention, an improved fish freshness fusion evaluation algorithm is used to perform feature-level fusion and time-series evolution analysis on characteristic spectral absorption bands, multi-level entropy sets, and indicative odor fingerprints. This includes mapping the characteristic spectral absorption bands to continuous numerical sequences characterizing the degree of protein degradation; mapping the multi-level entropy sets to continuous numerical sequences characterizing the degree of tissue structure deterioration; and mapping the indicative odor fingerprints to continuous numerical sequences characterizing the microbial spoilage process. The continuous numerical sequences characterizing the degree of protein degradation, the degree of tissue structure deterioration, and the microbial spoilage process are then subjected to independent timestamp alignment and interpolation processing to form a time-synchronized fused feature time series. A sliding window analysis is performed on the fused feature time series. Within each window, a dynamic freshness metric is calculated based on the weighted decision fusion architecture of the improved fish freshness fusion evaluation algorithm, and the value of the dynamic freshness metric at the center of the window is output.

[0088] The improved fish freshness fusion assessment algorithm incorporates a dynamic weight allocation mechanism based on a time decay factor. This mechanism involves defining initial weights for continuous numerical sequences representing protein degradation, tissue structure deterioration, and microbial spoilage processes. Based on a biochemical kinetic model of fish spoilage, the time point and rate of change for each indicator represented by the continuous numerical sequence during spoilage are preset. During the sliding window analysis, a time decay factor is calculated based on the time difference between the current analysis time and the initial change time point of each indicator; this factor increases with the time difference. The calculated time decay factor is multiplied by the initial weights of each sequence to obtain the dynamic weights of each sequence in the fusion calculation at the current time.

[0089] In practical implementation, an improved fish freshness fusion assessment algorithm is used to perform feature-level fusion and time-series evolution analysis on characteristic spectral absorption bands, multi-level entropy sets, and indicative odor fingerprints. This process includes mapping characteristic spectral absorption bands into continuous numerical sequences characterizing the degree of protein degradation. The mapping operation is achieved by extracting the absorption intensity and waveform symmetry of the characteristic spectral absorption bands and combining them into a scalar value. Mapping the multi-level entropy set into continuous numerical sequences characterizing the degree of tissue structure deterioration is achieved by calculating the weighted sum of entropy values ​​at all levels in the multi-level entropy set or selecting entropy values ​​at a specific scale. Mapping the indicative odor fingerprint into a continuous numerical sequence characterizing the microbial spoilage process is achieved by calculating the Euclidean or Mahalanobis distance between the indicative odor fingerprint vector and the odor fingerprint vector of a reference fresh sample.

[0090] For continuous numerical sequences characterizing the degree of protein degradation, the degree of tissue structural deterioration, and the microbial spoilage process, independent timestamp alignment and interpolation were performed. Timestamp alignment ensured that data sequences from different sources had a unified timeline. Interpolation methods, such as linear interpolation or spline interpolation, were used to supplement data at missing time points, thus forming a time-synchronized fused feature time series. A sliding window analysis was then performed on the fused feature time series. The sliding window had a preset window width and sliding step size. Within each window, a dynamic freshness metric was calculated based on a weighted decision fusion architecture using an improved fish freshness fusion assessment algorithm. The value of the dynamic freshness metric at the center of the window was then output.

[0091] In some embodiments, the specific method of feature mapping to a continuous numerical sequence can be chosen based on the properties of the feature itself. For example, mapping a feature spectral absorption band can use its absorption intensity value; mapping a multi-level entropy set can use the magnitude of its entropy gradient vector; and mapping an indicative odor fingerprint can use the first principal component score of its principal component score vector. Timestamp alignment is performed based on the physical time of data acquisition, and interpolation generates uniformly spaced data points at non-uniform sampling time points. The parameter settings for sliding window analysis can be determined based on the actual sampling frequency and the timescale of the spoilage process. The window width needs to cover a sufficient time range to reflect the changing trend, and the sliding step size determines the temporal resolution of the dynamic freshness metric output. The calculation of the dynamic freshness metric within the window is achieved by using a weighted decision fusion architecture to comprehensively calculate a scalar result from three sets of time-synchronized feature values ​​within the window.

[0092] The improved fish freshness fusion assessment algorithm incorporates a dynamic weight allocation mechanism based on a time decay factor. This mechanism involves defining initial weights for continuous numerical sequences representing protein degradation, tissue structure deterioration, and microbial spoilage processes. Based on a biochemical kinetic model of fish spoilage, the time point and rate of change for each indicator represented by the continuous numerical sequence during spoilage are preset. During the sliding window analysis, a time decay factor is calculated based on the time difference between the current analysis time and the initial change time point of each indicator; the time decay factor increases with the time difference. Multiplying the time decay factor by the initial weights of each sequence yields the dynamic weights of each sequence in the fusion calculation at the current time. Optionally, the time decay factor can be calculated using an exponential decay or linear growth model. An exemplary formula for calculating the time decay factor is as follows:

[0093] ;

[0094] in: Indicates the time of analysis At that time, the first A feature sequence (e.g., Represents protein degradation sequence, This represents the sequence of organizational structure deterioration. The time decay factor corresponding to the microbial spoilage sequence. Indicates the first The rate of change parameter of a characteristic sequence, derived from a biochemical kinetic model, is positive and its magnitude reflects the sensitivity of the indicator to changes in the process of decay. Indicates the first The starting point of change of the indicators represented by each feature sequence also originates from the biochemical kinetic model. This indicates the analysis time corresponding to the center of the current sliding window. This represents the time difference between the current moment and the point in time when the indicator began to change. Time decay factor It can be set to a baseline value (e.g., 1). Dynamic weights From initial weights With time decay factor Multiplying them together yields the result, i.e. The dynamic weights of each feature sequence are used to weight their corresponding feature values ​​when calculating the dynamic freshness quantification index. In practice, the initial change time points and change rate parameters of each index preset in the biochemical kinetic model can be determined based on prior knowledge or historical experimental data. See Table 1:

[0095] Table 1: Feature Sequence Parameters Used for Dynamic Weight Allocation

[0096] It is understandable that mapping different features into a unified continuous numerical sequence is the foundation for realizing time series fusion analysis. Timestamp alignment and interpolation solve the synchronization problem of multi-source heterogeneous data. Sliding window analysis provides a framework for local feature fusion and trend calculation in the time dimension. Based on the dynamic weight allocation mechanism of the time decay factor, by introducing the initial change time point and change rate parameters, the dependence of different indicators on different stages of spoilage can be dynamically adjusted. For example, in the early stage of spoilage, the weight of the tissue structure deterioration sequence may be enhanced by the time decay factor due to earlier changes, thus more accurately reflecting the freshness status. The entire implementation process, from feature mapping, data synchronization, window analysis to dynamic weighted fusion, constitutes the core computational steps for dynamic, multi-dimensional quantitative assessment of the fish spoilage process.

[0097] In one embodiment of the present invention, the adaptive correction process based on feature credibility in the improved fish freshness fusion evaluation algorithm includes the following steps: For a feature spectral absorption band, calculate its matching degree with a pre-stored standard database, and normalize the matching degree to the feature credibility of the feature spectral absorption band. For a multi-level entropy set, calculate its entropy value variation range based on historical data, and calculate the deviation score according to the current entropy value's position within this variation range, normalizing the deviation score to the feature credibility of the multi-level entropy set. For an indicative odor fingerprint, compare its principal component score with the Euclidean distance of the center point of a typical spoilage stage, and normalize the reciprocal of this Euclidean distance to the feature credibility of the indicative odor fingerprint. Based on the obtained feature credibility of the feature spectral absorption band, the feature credibility of the multi-level entropy set, and the feature credibility of the indicative odor fingerprint, dynamically adjust the corresponding fusion weight coefficients in the weighted decision fusion architecture. Using the adjusted fusion weight coefficients, perform weighted fusion calculation on the input feature spectral absorption band, multi-level entropy set, and indicative odor fingerprint.

[0098] For the characteristic spectral absorption band, its matching degree with a pre-stored standard database is calculated. This process includes retrieving the average spectral absorption band and its standard deviation band of standard fish samples at different freshness levels under the same characteristic absorption band from the pre-stored standard database. The waveform similarity measurement is then performed between the characteristic spectral absorption band to be tested and the standard spectral absorption band of each freshness level. This waveform similarity measurement includes correlation coefficient calculation and root mean square error calculation. Based on the calculated correlation coefficient and root mean square error, combined with the standard deviation band corresponding to each freshness level, the confidence probability that the characteristic spectral absorption band to be tested falls within the standard distribution of each freshness level is calculated. The confidence probability value corresponding to the freshness level with the highest confidence probability is selected as the matching degree between the characteristic spectral absorption band and the pre-stored standard database.

[0099] In its implementation, the improved fish freshness fusion evaluation algorithm incorporates an adaptive correction process based on feature reliability. This process includes: calculating the matching degree between the feature spectral absorption band and a pre-stored standard database, and normalizing the matching degree to the feature reliability of the feature spectral absorption band. Normalization maps the matching degree to a range of 0 to 1. For the multi-level entropy set, the entropy value variation range is calculated based on historical data. This historical data covers the entropy value set of samples throughout the entire process from fresh to spoilage. A deviation score is calculated based on the current entropy value's position within this variation range. This deviation score is obtained by comparing the current entropy value with the mean and standard deviation of the entropy value distribution under the corresponding freshness level in the historical data. The deviation score is then normalized to the feature reliability of the multi-level entropy set. Finally, for the indicative odor fingerprint, the principal component scores of the indicative odor fingerprint are compared with the Euclidean distance to the center points of typical spoilage stages. The Euclidean distance is calculated to represent the distance from the indicative odor fingerprint to the cluster centers of each spoilage stage in the dimensionality-reduced feature subspace. The reciprocal of the Euclidean distance is then normalized to the feature reliability of the indicative odor fingerprint. Based on the feature confidence levels of the characteristic spectral absorption bands, the multi-level entropy set, and the indicative odor fingerprint, the corresponding fusion weight coefficients in the weighted decision fusion architecture are dynamically adjusted. Using the adjusted fusion weight coefficients, a weighted fusion calculation is performed on the input characteristic spectral absorption bands, multi-level entropy set, and indicative odor fingerprint.

[0100] In some embodiments, the normalization of matching degree, deviation score, and inverse Euclidean distance can be achieved using methods such as max-min normalization or Z-score normalization followed by Sigmoid function transformation, ensuring that the final feature confidence value is between 0 and 1, with a larger value indicating higher confidence. Dynamic adjustment of the fusion weight coefficients can be achieved by multiplying or adding the feature confidence value with the baseline weights and then normalizing, ensuring that the sum of all adjusted weight coefficients is 1. Weighted fusion calculation can be performed by multiplying the numerical sequence mapped from each feature with its adjusted weight coefficients and then adding the results to obtain a comprehensive freshness evaluation value.

[0101] For the characteristic spectral absorption band, the matching degree between the characteristic spectral absorption band and a pre-stored standard database is calculated. This process includes retrieving the average spectral absorption band and its standard deviation band of standard fish samples at different freshness levels under the same characteristic absorption band from the pre-stored standard database. The waveform similarity of the characteristic spectral absorption band to be tested with the standard spectral absorption band of each freshness level is measured. This waveform similarity measurement includes correlation coefficient calculation and root mean square error (RMSE) calculation. Based on the correlation coefficient and RMS error, combined with the standard deviation band corresponding to each freshness level, the confidence probability that the characteristic spectral absorption band to be tested falls within the standard distribution of each freshness level is calculated. The confidence probability value corresponding to the freshness level with the highest confidence probability is selected as the matching degree between the characteristic spectral absorption band and the pre-stored standard database.

[0102] Optionally, the confidence probability can be calculated based on a multivariate Gaussian distribution model or through a joint probability model combining the correlation coefficient and the standardized root mean square error. A specific calculation example is as follows: Assume the characteristic spectral absorption band is composed of a vector X consisting of reflectance values ​​at n wavelengths, and the mean vector of the standard spectral absorption band for a certain freshness level i is... The covariance matrix is (This can be derived from its standard deviation band). Then, the confidence probability that the absorption band X of the measured characteristic spectrum belongs to level i is... It can be estimated using the following formula:

[0103] ;

[0104] in: This represents the confidence probability that the spectral absorption band X of the measured feature belongs to freshness level i. This represents the dimension of the characteristic spectral absorption band vector. This represents the vector of the characteristic spectral absorption band to be measured. This represents the average spectral absorption band vector of freshness level i in the pre-stored standard database. The covariance matrix representing the standard spectral absorption band vector of freshness level i reflects the standard deviation of reflectance at each wavelength point under that level and their interrelationships. Represents the covariance matrix The determinant of . Representing vectors The transpose of . Represents the covariance matrix The inverse matrix. It represents an exponential function, that is, an exponential operation with the natural constant e as the base. The waveform correlation coefficient represents the difference between the measured characteristic spectral absorption band X and the standard spectral absorption band μ_i of level i, used to supplement the measurement of shape similarity. Calculate the confidence probability for all levels. Then, take the maximum value. The degree of matching is represented by k, where k is the total number of freshness levels.

[0105] In practical implementation, the pre-stored standard database must contain a sufficient number of representative samples to construct a reliable statistical model. The correlation coefficient in waveform similarity measurement uses the Pearson correlation coefficient, and the root mean square error calculation requires standardization of the reflectance values ​​to eliminate the influence of dimensions. A simplified example of confidence probability calculation results is shown in Table 2.

[0106] Table 2: Confidence Probability Calculation Table for Matching Characteristic Spectral Absorption Bands with Freshness Grade Standards

[0107] Understandably, the adaptive correction process based on feature confidence dynamically adjusts the contribution of each input feature in the fusion decision by evaluating the reliability of each feature itself. The calculation of feature confidence compares the raw feature data with prior knowledge (pre-stored database, historical data distribution, typical cluster centers) to quantify its confidence level. For the matching degree calculation of feature spectral absorption bands, waveform similarity (correlation coefficient) and statistical distribution (multivariate Gaussian probability) are combined to provide a multi-dimensional similarity assessment. The probability value corresponding to the highest confidence level is used as the matching degree, reflecting the degree of fit between the tested feature and the best matching object in the standard library. Finally, the adjusted fusion weight coefficients give higher confidence features a greater weight in the final decision, while reducing the influence of lower confidence features, thereby improving the robustness of the fusion evaluation algorithm in the case of data quality fluctuations or partial feature failure. The weighted fusion calculation uses the adjusted weights to integrate multi-source information and output a comprehensive evaluation result.

[0108] In one embodiment of the present invention, the process of outputting a dynamic freshness metric and a shelf-life stage label for the fish meat to be tested includes defining the dynamic freshness metric as a continuous value between a preset minimum value and a preset maximum value, where the preset minimum value corresponds to a completely spoiled state and the preset maximum value corresponds to an absolutely fresh state. Multiple preset threshold intervals are defined, each corresponding to a shelf-life stage label, including fresh stage, quality decline stage, early spoilage stage, and inedible stage. The calculated dynamic freshness metric is compared with the preset threshold intervals to determine its corresponding threshold interval. The shelf-life stage label corresponding to the threshold interval is assigned to the fish meat to be tested. Finally, a freshness evaluation report is output, which includes the value of the dynamic freshness metric and the determined shelf-life stage label name.

[0109] In practice, the process outputs a dynamic freshness metric and a shelf-life stage label for the tested fish. This includes defining the dynamic freshness metric as a continuous value between a preset minimum and a preset maximum. The preset minimum corresponds to complete spoilage, and the preset maximum corresponds to absolute freshness. Multiple preset threshold ranges are defined, each corresponding to a shelf-life stage label, including freshness stage, quality decline stage, early spoilage stage, and inedible stage. The calculated dynamic freshness metric is compared with the preset threshold ranges. This comparison is achieved by determining which preset value range the dynamic freshness metric falls into, thus identifying the threshold range to which the dynamic freshness metric belongs. The shelf-life stage label corresponding to the threshold range is assigned to the tested fish. A freshness evaluation report containing the dynamic freshness metric value and the shelf-life stage label name is output.

[0110] In some embodiments, the preset minimum value of the dynamic freshness metric can be set to 0, and the preset maximum value can be set to 100, with higher values ​​indicating better freshness. The threshold range can be defined based on the statistical distribution of the dynamic freshness metric across a large number of historical samples. For example, the threshold range for the freshness stage is (80, 100], for the quality decline stage it is (60, 80], for the early spoilage stage it is (30, 60], and for the inedible stage it is [0, 30]. Comparing the dynamic freshness metric with the preset threshold range is a numerical range matching process, using sequential comparison or binary search algorithms to determine the range in which the metric value lies. Assigning a shelf-life stage label involves associating the text label corresponding to the matched range with the current sample. The freshness evaluation report can be output as a text file, a graphical interface display, or a database record. The report content must include at least the specific value of the dynamic freshness metric and the name of the shelf-life stage label. Optionally, the dynamic freshness metric F can be obtained by weighted fusion calculation of multi-source feature values ​​adjusted by time decay factors and feature confidence correction. A specific calculation example is as follows:

[0111] ;

[0112] in: This represents the dynamic freshness metric of the final output. This indicates the number of valid time points within the sliding window. This represents the total number of features involved in the fusion. In this scenario, M=3, which corresponds to the degree of protein degradation, the degree of tissue structure deterioration, and the microbial putrefaction process, respectively. This represents the time decay weighting coefficient at the j-th time point within the sliding window. This coefficient can be allocated based on how close this time point is to the current analysis time, in order to reflect the greater influence of recent data. This represents the fusion weight coefficient of the k-th feature after dynamic adjustment based on the time decay factor and feature credibility. This represents the feature confidence level of the k-th feature. This represents the feature value of the k-th feature at the j-th time point within the sliding window, after mapping and normalization. This feature value has been converted into a scalar positively correlated with freshness. The calculated dynamic freshness metric F is a continuous scalar value. The calculated dynamic freshness metric F is compared with a preset threshold interval. If the value of F is 85, the preset freshness stage threshold interval is (80, 100], and the quality decline stage threshold interval is (60, 80], then the F value of 85 is determined to belong to the (80, 100] interval, and therefore its threshold interval is determined to be the interval corresponding to the freshness stage. The shelf life stage label "freshness stage" is assigned to the fish sample being analyzed. An example of the final output freshness evaluation report content is: "Dynamic Freshness Metric: 85; Shelf Life Stage Label: Freshness Stage".

[0113] It is understandable that defining the numerical range and state correspondence of dynamic freshness metrics provides a unified and quantifiable benchmark for freshness evaluation. Pre-setting threshold intervals and corresponding labels discretizes continuous quantitative indicators into quality stages with clear physical meaning, facilitating understanding and decision-making. The process of comparing and determining threshold intervals is a crucial step in mapping the calculated abstract values ​​to specific quality levels. Assigning labels is the final operation to complete the mapping. The output report is the final presentation of the method, integrating quantitative evaluation results and qualitative classification conclusions with structured information. The calculation formula F of the dynamic freshness metrics integrates multi-feature information from multiple time points and weights it using weighting coefficients and reliability coefficients, reflecting the relative importance and reliability of different features at different time points, ultimately merging them into a comprehensive evaluation value.

[0114] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for evaluating the freshness of fish meat, characterized in that, The method includes: Acquire initial multi-source sensor data of the fish meat to be tested, including hyperspectral reflectance curve, surface texture microscopic image and volatile odor substance concentration spectrum; Wavelength-reflectance correlation analysis was performed on the hyperspectral reflectance curve to extract the characteristic spectral absorption bands that characterize the degradation state of fish myofibril protein. Multi-scale texture decomposition was performed on the surface texture microscopic image to calculate a multi-level entropy value set characterizing the integrity of the microstructure of fish meat tissue. Principal component dimensionality reduction and pattern matching were performed on the concentration spectrum of the volatile odor substances to identify indicative odor fingerprints associated with microbial metabolic activities; The characteristic spectral absorption bands, the multi-level entropy value set, and the indicative odor fingerprint are input into the improved fish freshness fusion evaluation algorithm; The improved fish freshness fusion evaluation algorithm is used to perform feature-level fusion and time series evolution analysis on the characteristic spectral absorption band, the multi-level entropy value set and the indicative odor fingerprint, and outputs the dynamic freshness quantification index and shelf life stage label of the fish to be tested. The improved fish freshness fusion evaluation algorithm adopts a weighted decision fusion architecture. The improvement is achieved by introducing a dynamic weight allocation mechanism based on time decay factor and an adaptive correction process based on feature credibility.

2. The method for evaluating the freshness of fish meat according to claim 1, characterized in that, Wavelength-reflectance correlation analysis was performed on the hyperspectral reflectance curves to extract characteristic spectral absorption bands characterizing the degradation state of fish myofibril proteins, including: The hyperspectral reflectance curve is smoothed and denoised to eliminate abnormal fluctuations in reflectance caused by equipment noise and ambient stray light. On the smoothed and denoised hyperspectral reflectance curve, the first derivative spectrum of each wavelength point is calculated, and the wavelength position where the reflectance change rate reaches a local maximum is identified and marked as a candidate absorption peak. Based on the fish biochemical component database, candidate absorption peaks located in the characteristic absorption bands of myofibril proteins are screened from the candidate absorption peaks to form a preliminary set of spectral absorption bands. Calculate the absorption intensity, full width at half maximum (FWHM), and waveform symmetry of each absorption band in the preliminary spectral absorption band set; The absorption intensity, full width at half maximum (FWHM), and waveform symmetry are compared with the pre-stored standard fresh fish spectral absorption band characteristics. The absorption bands with similarity exceeding the threshold are identified as characteristic spectral absorption bands representing the degradation state of myofibril proteins.

3. The method for evaluating the freshness of fish meat according to claim 1, characterized in that, Multi-scale texture decomposition is performed on the surface texture microscopic image to calculate a multi-level entropy value set characterizing the integrity of the microstructure of fish flesh tissue, including: The surface texture microscopic image is converted to grayscale and the region of interest containing fish muscle fiber bundles and connective tissue is segmented. Wavelet transform is used to decompose the image of the region of interest into multiple scales to obtain sub-band images containing horizontal, vertical and diagonal directional information at different scales; Calculate the energy entropy of each sub-band image at each scale, whereby the energy entropy reflects the complexity and irregularity of the image texture at each scale; The energy entropy calculated at all scales is arranged and combined in order from coarse to fine scale to generate the multi-level entropy value set. The gradient of entropy values ​​as a function of scale is extracted from the multi-level entropy value set. This gradient is used to quantify the degree of degradation of the microstructure of fish meat at different scales.

4. The method for evaluating the freshness of fish meat according to claim 1, characterized in that, Principal component dimension reduction and pattern matching were performed on the concentration spectrum of the volatile odor substances to identify indicative odor fingerprints associated with microbial metabolic activities, including: The concentration spectrum of the volatile odor substances is standardized and preprocessed to eliminate the influence of differences in the concentration dimensions between different odor substances; Principal component analysis was applied to the concentration spectrum of volatile odor substances after standardization pretreatment to calculate the eigenvalues ​​and eigenvectors of its covariance matrix; Based on the preset cumulative variance contribution rate threshold, the number of principal components is selected, and the original high-dimensional odor concentration data is projected onto the feature subspace composed of the selected principal components to obtain the dimensionality-reduced principal component score data. Using a pre-established fish meat spoilage stage odor database, which contains clustering center patterns of samples at different spoilage stages in the principal component feature subspace; Calculate the similarity distance between the dimensionality-reduced principal component score data and the cluster center patterns of each spoilage stage in the fish spoilage stage odor database; The feature vectors corresponding to the putrefaction stage cluster centers with the smallest similarity distance to the reduced principal component score data are combined to determine the indicative odor fingerprint that characterizes the current sample's microbial metabolic activity.

5. The method for evaluating the freshness of fish meat according to claim 1, characterized in that, The improved fish freshness fusion evaluation algorithm is used to perform feature-level fusion and time-series evolution analysis on the characteristic spectral absorption bands, the multi-level entropy value set, and the indicative odor fingerprint, including: The characteristic spectral absorption bands are mapped into a continuous numerical sequence characterizing the degree of protein degradation; The multi-level entropy value set is mapped into a continuous numerical sequence characterizing the degree of tissue structure deterioration; The indicative odor fingerprint is mapped into a continuous numerical sequence characterizing the microbial putrefaction process; The continuous numerical sequences characterizing the degree of protein degradation, the continuous numerical sequences characterizing the degree of tissue structural deterioration, and the continuous numerical sequences characterizing the microbial putrefaction process are each subjected to independent timestamp alignment and interpolation processing to form a time-synchronized fused feature time series. A sliding window analysis is performed on the fusion feature time series. Within each window, a dynamic freshness metric is calculated based on the weighted decision fusion architecture of the improved fish freshness fusion evaluation algorithm. The value of the dynamic freshness metric at the center of the window is then output.

6. The method for evaluating the freshness of fish meat according to claim 5, characterized in that, The dynamic weight allocation mechanism based on the time decay factor in the improved fish freshness fusion evaluation algorithm includes: An initial weight is defined for the continuous numerical sequence characterizing the degree of protein degradation, the continuous numerical sequence characterizing the degree of tissue structural deterioration, and the continuous numerical sequence characterizing the microbial putrefaction process, respectively. Based on the biochemical dynamics model of fish spoilage, the time point and rate of change of the indicator represented by each continuous numerical sequence are preset during the spoilage process. During the sliding window analysis, a time decay factor is calculated based on the time difference between the current analysis time and the starting time of each indicator. The time decay factor increases as the time difference increases. The time decay factor is multiplied by the initial weight of each sequence to obtain the dynamic weight of each sequence in the fusion calculation at the current time.

7. The method for evaluating the freshness of fish meat according to claim 1, characterized in that, The adaptive correction process based on feature reliability in the improved fish freshness fusion evaluation algorithm includes: For the characteristic spectral absorption band, calculate its matching degree with the pre-stored standard database, and normalize the matching degree to the characteristic confidence degree of the characteristic spectral absorption band; For the multi-level entropy value set, the range of entropy value change is calculated based on historical data, and the deviation score is calculated according to the position of the current entropy value within the range of change. The deviation score is then normalized to the feature credibility of the multi-level entropy value set. For the indicative odor fingerprint, its principal component score is compared with the Euclidean distance of the center point of the typical putrefaction stage, and the reciprocal of the Euclidean distance is normalized to the feature confidence of the indicative odor fingerprint. Based on the feature confidence of the characteristic spectral absorption band, the feature confidence of the multi-level entropy value set, and the feature confidence of the indicative odor fingerprint, the corresponding fusion weight coefficients in the weighted decision fusion architecture are dynamically adjusted. Using the adjusted fusion weight coefficients, a weighted fusion calculation is performed on the input characteristic spectral absorption bands, the multi-level entropy value set, and the indicative odor fingerprint.

8. The method for evaluating the freshness of fish meat according to claim 7, characterized in that, The step of calculating the matching degree between the characteristic spectral absorption band and the pre-stored standard database includes: From the pre-stored standard database, retrieve the average spectral absorption band and its standard deviation band of standard fish meat samples of different freshness grades under the same characteristic absorption band; The waveform similarity of the characteristic spectral absorption band to be tested with the standard spectral absorption band of each freshness grade is measured. The waveform similarity measurement includes correlation coefficient calculation and root mean square error calculation. Based on the correlation coefficient and the root mean square error, and combined with the standard deviation band corresponding to the freshness grade, calculate the confidence probability that the absorption band of the spectral feature to be tested falls into the standard distribution of the freshness grade. The confidence probability value corresponding to the freshness level with the highest confidence probability is selected as the matching degree between the characteristic spectral absorption band and the pre-stored standard database.

9. The method for evaluating the freshness of fish meat according to claim 1, characterized in that, The output of the dynamic freshness quantification indicators and shelf-life stage labels of the fish meat to be tested includes: The dynamic freshness metric is a continuous value between a preset minimum value and a preset maximum value, where the preset minimum value corresponds to a completely rotten state and the preset maximum value corresponds to an absolutely fresh state. Multiple threshold ranges are preset, and each threshold range corresponds to a shelf life stage label. The shelf life stage label includes the fresh stage, the quality decline stage, the early spoilage stage, and the inedible stage. The calculated dynamic freshness metric is compared with a preset threshold range to determine the threshold range to which it belongs. Assign the shelf life stage label corresponding to the threshold interval to the fish meat to be tested; The output includes a freshness evaluation report containing the numerical values ​​of the dynamic freshness metric and the label name of the shelf life stage.

10. A fish meat freshness evaluation system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the fish meat freshness evaluation method according to any one of claims 1 to 9.

Images