Food taste sensory data image fingerprint construction method and food taste recognition method

Abstract

The application discloses a food taste sensory data image fingerprint construction method and a food taste recognition method, and comprises the following steps: collecting one-dimensional time sequence response data of a food sample under a taste detection system, and obtaining a standardized response sequence through pretreatment; performing event boundary detection based on local mutation intensity, cutting the whole response into multiple sub-events, and resampling and aligning; extracting morphological, spectral, cepstrum, time-frequency energy, graph structure and sequence pattern complexity features of the sub-events to form an event feature matrix; constructing an event relationship matrix according to the amplitude, area, cepstrum and graph spectrum relationship between events; constructing a recursive texture matrix based on the centroid vector of each sub-event; and finally generating a three-channel visual taste fingerprint map according to the event feature matrix, the event relationship matrix and the recursive texture matrix. The application can fully mine deep dynamic information in the taste response, generate a taste fingerprint map with speciality and stability, and is suitable for food taste classification, quality evaluation and authenticity discrimination.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensory analysis technology for food, specifically to a method for constructing image fingerprints of food taste sensory data and a method for food taste recognition, and more particularly to a method for feature extraction, encoding representation and dedicated fingerprint construction of one-dimensional time-series response data and a method for food taste recognition. Background Technology

[0002] Food taste is one of the core pieces of information in food quality evaluation, flavor analysis, product grading, quality control, and intelligent manufacturing. Current food taste analysis mainly relies on human sensory evaluation, physicochemical testing, and taste system detection technologies. Human sensory evaluation is highly subjective, and its repeatability is significantly affected by the test subject's condition; while physicochemical analysis has high precision, it usually targets single components and cannot directly reflect the overall taste perception; the taste system, through electrochemical sensor arrays, can rapidly respond to the overall taste attributes of a sample, thus becoming an important direction for the digitalization of food sensory perception.

[0003] However, existing methods for processing taste system data still primarily rely on a limited number of explicit features, such as extreme points, inflection points, peak areas, and peak heights, in the original volt-ampere curves and current response curves. While these methods are simple to implement, they suffer from the following problems:

[0004] First, relying solely on a few manually defined points leads to the loss of a large amount of dynamic response information, which cannot fully characterize the complex taste differences of the samples.

[0005] Second, it is highly sensitive to noise, baseline drift, electrode state fluctuations, and sampling time stretching, resulting in insufficient repeatability.

[0006] Third, extracting features only from local points in the time domain lacks the ability to represent frequency domain, time-frequency domain, structural domain, and cross-event response relationships, making it difficult to construct stable, specific, and transferable taste fingerprints.

[0007] Fourth, most existing methods can only obtain a numerical feature vector, lacking a visual fingerprint representation that is easy for both humans and machines to use, which is not conducive to the construction of standardized databases and intelligent comparison.

[0008] Therefore, how to form a taste fingerprint that is both visual and machine-recognizable, and has high stability and unique recognition capabilities, so as to achieve accurate identification of food taste, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] In view of the above problems, this invention is proposed to provide a method for constructing image fingerprints of food taste sensory data and a method for food taste recognition that overcomes or at least partially solves the above problems. It can break through the traditional approach of extracting extreme points and inflection points, and systematically encode the dynamic morphological information, spectral information, cepstral information, graph structure information and inter-event coupling relationships contained in the one-dimensional time-series response of the taste system. Finally, it forms a taste sensory data fingerprint that is both visual and machine-recognizable, and has specificity and stability, for use in the accurate classification, quality evaluation, authenticity identification and batch tracking of food taste.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a method for constructing image fingerprints of food taste sensory data, comprising the following steps: S1: Collect one-dimensional time-series response data of food samples under the taste detection system; S2: Preprocess the one-dimensional time-series response data to obtain a standardized response sequence; S3: Based on the local mutation intensity of the standardized response sequence, perform event boundary detection, divide the entire response sequence into multiple response sub-events based on the event boundary points, and perform uniform length resampling on each response sub-event to obtain standardized response sub-events; S4: Extract multi-domain features for each standardized response sub-event, including at least morphological features, spectral features, cepstral features, time-frequency energy features, graph structure features, and order pattern complexity features, to form an event feature matrix; S5: Construct an event relationship matrix based on the amplitude relationship, area relationship, cepstral relationship, and spectral relationship between different events in the event feature matrix; S6: Construct a recursive texture matrix based on the centroid vectors of each standardized response sub-event; S7: Generate a three-channel visual taste fingerprint based on the event feature matrix, the event relationship matrix, and the recursive texture matrix.

[0011] Preferably, the one-dimensional time-series response data is a current response sequence, potential response sequence, or conductivity response sequence collected during the steady-state stage of the reference liquid, the sample stimulation stage, and the cleaning and regeneration detection stage.

[0012] Preferably, the event boundary detection step in step S3 includes: calculating the first-order difference and second-order difference of the sequence amplitude corresponding to the standardized response sub-event; fusing the first-order difference and second-order difference according to a preset weight to obtain a fusion energy function; selecting local peak points in the fusion energy function that exceed a predetermined multiple of the mean standard deviation as candidate boundary points; and applying a minimum time interval constraint to obtain an effective boundary set.

[0013] Preferably, in the event feature matrix of step S4: The morphological features include at least positive peak value, negative peak value, peak spacing, rise time constant, recovery time constant, signed area, curvature energy, and zero crossover delay; The spectral features include at least the spectral centroid, spectral bandwidth, spectral roll-off point, spectral flatness, spectral entropy, spectral flux, dominant modulation frequency, and high-to-low frequency energy ratio. The cepstral features are obtained by using electrochemical cepstral coefficients, which are obtained by integrating the amplitude spectrum of the response sub-events with a non-uniform filter bank, logarithmic compression, and discrete cosine transform. The time-frequency energy features are decomposed using a three-layer wavelet packet decomposition. The energy of each sub-band node is normalized to obtain the wavelet packet energy distribution vector. The graph structure features are obtained by converting one-dimensional response sub-events into an undirected graph using the visual graph method, and at least one of the following is extracted as graph structure features: average degree, clustering coefficient, graph radius, graph entropy, degree distribution skewness, and average path length. The order pattern complexity features include at least permutation entropy, order transition entropy, and statistical complexity.

[0014] Preferably, the elements of the event relationship matrix in step S5 are obtained by weighting the amplitude ratio, area ratio, cepstral angle, and spectral distance between any two standardized response sub-events according to preset weights.

[0015] Preferably, the recursive texture matrix construction step in step S6 includes: S61: For each standardized response sub-event, calculate the normalized time centroid and amplitude centroid to form the event centroid vector; S62: Calculate the distance between the centroid vectors of any two events, and map the distance to a similarity to obtain a recursive texture matrix.

[0016] Preferably, step S7 includes: The event feature matrix is ​​arranged in the order of event number and event feature dimension. After normalization and grayscale quantization of the event features, it is mapped to a first channel grayscale image. The event relationship matrix is ​​arranged in the order of event number and relationship strength dimension. The relationship strength is normalized and grayscaled, and then mapped to a second channel grayscale image. The recursive texture matrix is ​​arranged according to the event sequence number and the recursive similarity dimension. The recursive similarity is normalized and grayscale is then mapped to a third-channel grayscale image. The first channel grayscale image, the second channel grayscale image, and the third channel grayscale image are scaled to the same size and then superimposed to obtain a three-channel visual taste fingerprint image.

[0017] Preferably, the method further includes step S8: quantizing the three-channel visual taste fingerprint to generate a binary taste fingerprint code.

[0018] Preferably, the step of generating the binary taste fingerprint code includes: S81: Perform grayscale fusion on the three-channel visualized taste fingerprint to obtain a fused grayscale image; S82: Perform a two-dimensional discrete cosine transform on the fused grayscale image, extract the low-frequency coefficient value region, remove the DC component from the low-frequency coefficient value region, and then calculate the median threshold; wherein, the low-frequency coefficient value region is the coefficient value region with a frequency value lower than the set threshold. S83: Binarize the corresponding two-dimensional discrete cosine transform results of each coordinate position of the fused grayscale image according to the median threshold to obtain the binary values ​​of each coordinate position, forming a binary sequence.

[0019] Secondly, embodiments of the present invention provide a food taste recognition method, which constructs a three-channel visualized taste fingerprint database for different foods using the food taste sensory data image fingerprint construction method described in the first aspect; including the following steps: The first aspect of the food taste sensory data image fingerprint construction method is used to construct the taste fingerprint of the food sample to be tested. The taste fingerprint of the food sample to be tested for comprehensive taste quality is compared with at least one taste fingerprint in the comprehensive taste quality three-channel visualized taste fingerprint database, and the distance between the two is calculated. Based on the relationship between the comprehensive taste distance and the preset threshold, the taste category or quality status of the food sample to be tested is determined.

[0020] The specific beneficial effects of the above-mentioned technical solutions provided in the embodiments of the present invention include at least the following: First, this invention is geared towards the detection scenario of the taste system. It is no longer limited to a few explicit features such as extreme points, inflection points, and peak areas. Instead, it extracts taste response information from six levels: morphological domain, spectral domain, cepstral domain, time-frequency domain, graph structure domain, and sequence mode domain. This allows for a more complete preservation of dynamic information during sample stimulation, interface polarization, and recovery decay.

[0021] Second, the present invention introduces an event boundary detection and segment alignment mechanism, which automatically decomposes the entire taste system response into multiple standardized response sub-events, which can better adapt to various detection procedures such as pulse excitation, step excitation, and cyclic excitation, and improve the consistency between repeated measurements.

[0022] Third, this invention proposes electrochemical cepstral coefficients (ECC), which transfers the cepstral concept from the field of speech recognition to the analysis of the taste system response. It also reconstructs the filter bank by combining the transient response characteristics of the taste system, thereby capturing the hidden spectral structure and dynamic recovery mode that are difficult to characterize by the traditional peak method.

[0023] Fourth, this invention proposes a graph structure feature extraction method based on the visual graph method, which converts the one-dimensional taste response into a graph topology structure. The morphological organization of the taste response is described by parameters such as average degree, clustering coefficient, spectral radius and graph entropy, so that the taste signal can obtain structural hierarchical representation ability.

[0024] Fifth, by constructing an event relationship matrix and a recursive texture matrix, this invention incorporates cross-event coupling relationships, repetitive structures, and relative strength relationships into the taste fingerprint, which can better reflect the dynamic organizational differences under different taste stimuli such as sour, sweet, bitter, salty, umami, astringent, numbing, and spicy, as well as comprehensive taste stimuli.

[0025] Sixth, this invention simultaneously outputs a three-channel visual taste fingerprint map and a binary fingerprint code, enabling the results to be displayed intuitively and quickly indexed and compared by machines, facilitating the formation of a standardized food taste sensory database.

[0026] Seventh, this invention has strong versatility and is applicable to the analysis of the taste attributes of foods such as sour, sweet, bitter, salty, umami, numbing, and spicy tastes. It is also applicable to the overall taste characterization of alcoholic beverages, tea drinks, fermented products, condiments, plant extracts, and compound food systems. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the overall process of a method for constructing a sensory fingerprint of food taste data according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the preprocessing and event boundary detection process for one-dimensional taste temporal response data provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the extraction process of electrochemical cepstral coefficients (ECC) provided in an embodiment of the present invention; Figure 4 A schematic diagram of the graph structure feature construction process based on the visual chart method provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of uniform-length resampling and segment alignment based on actual taste system sample data provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the event matrix after resampling based on actual taste system sample data provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of electrochemical cepstral feature (ECC) extraction based on actual taste system sample data provided in an embodiment of the present invention. Figure 8 This is a schematic diagram illustrating the construction of graph structure features based on actual taste system sample data provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of a three-channel taste fingerprint based on actual taste system sample data provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the three-channel taste fingerprint image and binary fingerprint code generation process provided in an embodiment of the present invention; Figure 11 Binary fingerprint code based on actual taste system sample data provided in embodiments of the present invention H Generate a schematic diagram. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention specifically discloses a method for constructing image fingerprints of food taste sensory data, the process of which is as follows: Figure 1 As shown, the method steps are as follows: S1: Collect one-dimensional time-series response data of food samples under the taste detection system; S2: Preprocess the one-dimensional time-series response data to obtain a standardized response sequence; S3: Event boundary detection is performed based on the local mutation intensity of the standardized response sequence. The entire response sequence is divided into multiple response sub-events based on the event boundary points, and each response sub-event is resampled to a uniform length to obtain the standardized response sub-events. S4: Extract multi-domain features for each standardized response sub-event, including at least morphological features, spectral features, cepstral features, time-frequency energy features, graph structure features, and order pattern complexity features, to form an event feature matrix; S5: Construct an event relationship matrix based on the amplitude relationship, area relationship, cepstral relationship and spectral relationship between different events in the event feature matrix; S6: Construct a recursive texture matrix based on the centroid vectors of each standardized response sub-event; S7: Generate a three-channel visual taste fingerprint based on the event feature matrix, event relationship matrix, and recursive texture matrix.

[0031] In one embodiment, the taste detection system in step S1 is a voltammetric taste system, which can be an electronic tongue detection system composed of cross-sensitive taste sensors or an electrochemical electrode array, collecting at least one one-dimensional time-series response signal generated by the food sample under the following cycles: steady state in reference solution, sample stimulation, and washing, regeneration, and detection. ,in This is the sampling point number.

[0032] In this embodiment, the one-dimensional time-series response data are current response sequences, potential response sequences, or conductivity response sequences collected during the steady-state stage of the reference liquid, the sample stimulation stage, and the cleaning and regeneration detection stage.

[0033] In practical implementation, the voltammetric electronic tongue system includes a working electrode array, a reference electrode, a counter electrode, a constant potential circuit control unit, a temperature monitoring unit, and a sample cell. Each working electrode has cross-sensitivity to different flavor substances. The sample detection process follows a cyclical process of standard solution activation, sample injection, response acquisition, and cleaning solution regeneration to obtain continuous time-series response data that can reflect flavor stimulation, interface polarization, and recovery behavior.

[0034] In one embodiment, such as Figure 2 As shown, the preprocessing in step S2 includes smoothing and denoising, baseline correction, temperature compensation, and robust normalization.

[0035] In this embodiment, Savitzky-Golay filtering is used for smoothing and denoising. Specifically, the window length is 9 and the fitting order is 3.

[0036] In this embodiment, baseline correction employs the asymmetric least squares method. Specifically, the baseline is solved... Satisfy the objective function:

[0037] In the formula, Fit the objective function to the baseline. For the first The weighting coefficients for each sampling point As a smoothing constraint factor, This represents the second-order difference term of the baseline sequence. Weighting coefficients. Defined as:

[0038] in, This is the asymmetric weighting parameter, with a value of 0.001.

[0039] In this embodiment, temperature compensation employs linear correction to a reference temperature. Specifically, temperature compensation uses a reference temperature... After performing linear correction, the temperature-compensated sequence is obtained:

[0040] In the formula, This is the response value after temperature compensation. The temperature drift coefficient obtained from pre-calibration. The temperature at the time of detection.

[0041] In this embodiment, robust normalization is performed using the median and the median absolute difference (MAD) to obtain the standardized response sequence:

[0042] In the formula, To standardize the response sequence, To be the median of the compensated sequence, This represents the absolute median difference.

[0043] The above preprocessing examples are not only used to improve the signal-to-noise ratio, but also to eliminate electrode baseline drift, temperature drift effect and batch-to-batch background shift that are common in taste system detection, so that subsequent features are closer to the intrinsic taste response of food samples.

[0044] In one embodiment, the event boundary detection step in step S3 includes: normalizing the response sequence. Perform event boundary detection and segment alignment to obtain K Each response sub-event. Specifically, the first and second differences of the sequence amplitude corresponding to the standardized response sub-event are calculated. The first and second differences are fused according to preset weights to obtain a fused energy function. Local peak points in the fused energy function that exceed a predetermined multiple of the mean standard deviation are selected as candidate boundary points, and a minimum time interval constraint is applied to obtain the effective boundary set.

[0045] It should be noted that in taste detection scenarios, processes such as sample introduction, sensor interface adsorption, diffusion, polarization enhancement, peak response, and rebalancing can cause local abrupt changes, slope changes, or curvature changes in the one-dimensional response curve. This embodiment does not pre-identify the types of these micro-processes. Instead, it constructs a fusion energy function based on the first and second differences of the standardized response sequence, automatically detecting effective boundary points from signal change characteristics. The effective boundary set is used to determine the start and end points of each response sub-event, dividing the entire response sequence into multiple local dynamic response units. Each response sub-event can correspond to one or more of the aforementioned taste response micro-processes. Therefore, compared to directly extracting a single peak from the entire curve, the event-based segmentation method can more fully preserve the dynamic response process of the taste system to food flavor stimuli.

[0046] like Figure 2As shown, in actual execution, the absolute value of the first-order difference is first calculated:

[0047] Next, calculate the absolute value of the second difference:

[0048] Based on this, a fusion energy function is constructed:

[0049] in, Used to characterize the intensity of local abrupt changes in the response curve. When the following equation is satisfied, the corresponding peak point is identified as a candidate boundary point:

[0050] In the formula, The mean of the fusion energy function, This represents the standard deviation of the fusion energy function. This is then combined with the minimum boundary margin constraint. By eliminating duplicate trigger points, a set of valid boundaries is obtained, and the entire sequence is divided into segments. A response sub-event In this embodiment, each response sub-event is uniformly mapped to... One sampling point, preferred This is to mitigate the impact of differences in sampling time scaling.

[0051] In one embodiment, the event feature matrix in step S4 is used to characterize the taste response characteristics of the food sample from different levels, such as amplitude strength, response rhythm, recovery tail, local fluctuations, topological organization, and dynamic complexity, wherein: Morphological characteristics include at least a positive peak value. negative peak Peak spacing Rise time constant Recovery time constant Area with sign Curvature energy and zero cross-delay It should be noted that morphological characteristics are used to characterize the adsorption, diffusion, polarization, depolarization, and recovery processes of the taste sensing interface under sample stimulation. The marked area is defined as:

[0052] In the formula, Characterizes the cumulative response strength of a sub-event over the entire duration. The sampling time interval is denoted as . Curvature energy is defined as:

[0053] In the formula, This is used to characterize the curvature and local morphological complexity of sub-event curves. These morphological parameters can be used to represent the amplitude, duration, recovery trajectory, and local geometry of the taste response.

[0054] The positive peak value and the area under sign can reflect the overall response intensity during the stimulus phase, the rise time constant and the recovery time constant can reflect the mass transfer and recovery speed of the taste sensor interface, and the zero cross-delay helps to characterize the response transition behavior formed by inhibition-recovery or enhancement-attenuation.

[0055] Spectral features are extracted through frequency domain analysis of each response sub-event, including at least the spectral centroid, spectral bandwidth, spectral roll-off point, spectral flatness, spectral entropy, spectral flux, dominant modulation frequency, and high-to-low frequency energy ratio. It should be noted that these spectral features are used to characterize the rhythmic changes in taste response, energy distribution, and dynamic stability of the sensor array under different stimuli from various flavoring substances. Among these, the spectral centroid... The centroid location used to characterize the response spectrum is defined as:

[0056] In the formula, For frequency The power spectral density or spectral energy at that point. The larger the spectral centroid, the higher the proportion of high-frequency components in the response event, and the faster its transient changes.

[0057] Spectral entropy The degree of dispersion and uncertainty in the spectral energy distribution is defined as:

[0058] in,

[0059] In the formula, This represents the normalized spectral probability distribution. A higher spectral entropy indicates a more dispersed spectral distribution and a more complex response structure.

[0060] In practical applications, the spectral centroid can reflect the activity level of fast-changing components in the taste response, while the spectral entropy can characterize the degree of mixing of multiple frequency components. Combining the two can effectively compensate for the shortcomings of the traditional peak method in characterizing the frequency domain structure. For samples with fast responses such as sour and salty tastes, as well as samples with obvious tailing and slow-release characteristics such as bitter and astringent tastes, their spectral distribution and energy transfer patterns often show significant differences.

[0061] Cepstral characteristics were extracted using an electrochemical cepstral coefficient extraction method. For example... Figure 3 As shown, this method first performs a Fast Fourier Transform on the response sub-events to obtain the amplitude spectrum. Then, by integrating the spectral energy through a non-uniform triangular filter bank, the energy vector of the filter bank is obtained. The logarithm of the coefficients was then taken and a discrete cosine transform was performed to extract the first 12 dimensions of cepstral coefficients, defined as electrochemical cepstral coefficients (ECC). This feature draws inspiration from cepstral analysis in the audio field, but the center frequency and bandwidth of the filter bank were reconstructed to suit the transient response of the taste system. It can characterize hidden periodicity, recovery trends, and weak oscillations that are difficult to express using traditional peak methods. Compared to directly using the original spectrum, electrochemical cepstral coefficients are better suited for compressively expressing the slow recovery trends, periodic perturbations, and subtle differences in the sensor array in the taste system response, thus improving the ability to distinguish subtle taste differences in complex food systems.

[0062] The time-frequency energy characteristics are implemented using a three-level wavelet packet decomposition. The db4 wavelet basis is selected, and the energy of each sub-band node is normalized to obtain the wavelet packet energy distribution vector.

[0063] In the formula, to These correspond to the energy proportions at different time scales and frequency bands. This vector is used to characterize the energy distribution pattern of the taste response at different time scales.

[0064] The graph structure feature uses the visual graph method to convert one-dimensional response sub-events into an undirected graph. If two sampling points... and Satisfy any intermediate point All of the following are available:

[0065] in, For nodes One-dimensional time-series response data values, For nodes The one-dimensional time-series response data value is then considered to be the node. With nodes Visual connections are established, and edges are created in the graph. For example... Figure 4 As shown, further extraction of average degree Clustering coefficient Spectral radius Graph entropy skewness of degree distribution and average path length As a graph structural feature, by mapping time-series curves to a graph topology, the morphological complexity and global relationships of taste responses can be described at the structural organization level. This type of graph structural feature can transform the peak-valley connectivity, local clustering degree, and overall structural order of taste response curves into quantifiable parameters, thereby enhancing the ability to identify complex complex tastes and subtle differences between similar samples.

[0066] The complexity of the order pattern is characterized by permutation entropy, order transition entropy, and statistical complexity. An embedding dimension is set. ,Delay Construct an ordered pattern distribution for the response sub-events, and arrange the entropy. Defined as:

[0067] In the formula, For the first The probability of the occurrence of a seed sequence pattern. This represents the total number of order patterns. A higher permutation entropy indicates a more disordered and complex response sequence. Further, order transition entropy and statistical complexity are calculated based on the adjacent order pattern transition probability matrix to describe local dynamic organizational characteristics and microstructural variability. When multiple flavor compounds in a food sample act together, the response sequence often exhibits higher order pattern transition complexity; therefore, order pattern complexity characteristics help characterize taste coupling behavior under combined flavor stimuli.

[0068] Through the above steps, a fixed-dimensional feature vector is obtained for each response sub-event. All K The response sub-events are arranged in chronological order to form an event feature matrix: .

[0069] In one embodiment, the elements of the event relationship matrix in step S5 are obtained by weighting the amplitude ratio, area ratio, cepstral angle, and spectral distance between any two standardized response sub-events according to preset weights.

[0070] In actual execution, for any two events and Calculate their amplitude ratios respectively. Area ratio cepstral angle and spectral distance Then, by weighting and combining them, we obtain the elements of the event relationship matrix:

[0071] in, 、 、 、 For fixed weights, values ​​of 0.30, 0.25, 0.25, and 0.20 are used respectively. This step explicitly encodes the coupling relationships between adjacent and distant events in the original one-dimensional time-series response, forming a relational structure with grammatical features of taste response. This relational structure not only reflects the strength differences of individual sub-events but also the synergistic relationships between the various stages of the taste response from stimulus, polarization to recovery, and can be used to express the overall taste organization patterns of the same food sample.

[0072] In one embodiment, the construction step of the recursive texture matrix in step S6 includes: S61: For each standardized response sub-event, calculate the normalized time centroid and amplitude centroid to form the event centroid vector; S62: Calculate the distance between the centroid vectors of any two events and map the distance to a similarity to obtain the recursive texture matrix.

[0073] In practice, the recursive relationships between the centroid vectors of each event are extracted to construct a recursive texture matrix. P The process includes: for the first i A standardized response sub-event after uniform length resampling:

[0074] in, Indicates the first The amplitude sequence obtained by resampling a standardized response sub-event with a uniform length. , … These respectively indicate that the response sub-event is in the 1st, 2nd, ..., 1st... The standardized response amplitude at each resampling location.

[0075] Specifically, the entire one-dimensional time-series response data is preprocessed to obtain a standardized response sequence, which is then segmented into multiple response sub-events using event boundary detection. Since the original duration and number of sampling points of different response sub-events may differ, each response sub-event is resampled to a uniform length, so that they are all mapped to a length of... Equal-length sequences.

[0076] The standardized response amplitude can be a standardized current response value, a standardized potential response value, or a standardized conductivity response value, depending on the type of data acquired by the original taste detection system. If the original one-dimensional time-series response data is a current response sequence, then... This represents the standardized current response amplitude after preprocessing, segmentation, and uniform-length resampling; if the original data is a potential response sequence or a conductance response sequence, then... These represent the corresponding normalized potential response amplitude or normalized conductivity response amplitude, respectively. The purpose of this expression is to uniformly represent each response sub-event as an amplitude vector of consistent length, facilitating subsequent calculations of the event centroid vector, recursive texture matrix, and three-channel taste fingerprint.

[0077] It should be noted that each response sub-event, after being resampled to a uniform length, becomes a length of [length missing]. sequence.

[0078] First As the first m Response weights for each sampling point:

[0079] in, Indicates the sequence number of the sub-event being responded to. This represents the total number of response sub-events obtained after event boundary detection; Indicates the first The sequence number of the sampling points after uniform-length resampling of each response sub-event. The fixed number of points after uniform length resampling; express The absolute value is used to characterize the magnitude of the response intensity at that sampling point. The reason for using the absolute value is that the taste system response may contain both positive and negative responses. If the signed amplitude is used directly to calculate the weight, the positive and negative responses may cancel each other out. By using the absolute value, both positive and negative peaks can be included in the weight calculation according to their response intensity. This represents a very small positive number, used to prevent the weight from being zero when the amplitude at certain sampling points is zero, and also to avoid the denominator being zero in subsequent centroid calculations. This parameter is only used for numerical stability and can generally be set to a value of [value missing]. , Or other positive numbers that are much smaller than the normal response amplitude will not substantially change the relative size of the response weight; Indicates the first The first standardized response sub-event The response weights of each resampled point are used to ensure that sampling points with larger response amplitudes contribute more to the subsequent calculation of the time centroid and amplitude centroid, while sampling points with smaller response amplitudes contribute less.

[0080] Therefore, the purpose of this formula is to assign weights to each resampling point based on the magnitude of the response amplitude, for subsequent calculation of the normalized time centroid and amplitude centroid. This weighting design highlights the main peaks, valleys, or regions of significant dynamic change in the response sub-event, making the event centroid vector more accurately reflect the main response location and amplitude center of the sub-event.

[0081] Calculate its normalized time centroid :

[0082] Calculate its amplitude centroid :

[0083] And form the centroid vector of the event :

[0084] in, This is used to characterize the distribution of the main response energy of the sub-event on the normalized time axis. Used to characterize the amplitude center of the response sub-event. For any two response sub-events and Calculate the distance between its centroid vectors. :

[0085] Then, the distance is converted into similarity using an exponential kernel function, resulting in the elements of the recursive texture matrix. :

[0086] Matrix elements The larger the value, the closer the principal response positions and amplitude centers of the two sub-events are; matrix elements The smaller the value, the greater the difference between the two. The resulting recursive texture matrix is ​​used to characterize the repetitive patterns and rhythmic structure across events in the entire taste response, and serves as the third channel of the three-channel taste fingerprint.

[0087] In one embodiment, step S7 includes: The comprehensive flavor event feature matrix is ​​arranged in order of event number and event feature dimension. After normalization and grayscale quantization of the event features, they are mapped to a first-channel grayscale image. ; The comprehensive flavor event relationship matrix is ​​arranged in order of event number and relationship strength dimension. After normalizing and gray-quantifying the relationship strength, it is mapped to a second-channel grayscale image. ; The comprehensive flavor recursive texture matrix is ​​arranged according to the event sequence number and the recursive similarity dimension. After normalizing and grayscale quantization of the recursive similarity, it is mapped to a third-channel grayscale image. ; The first channel grayscale image Second channel grayscale image and the third channel grayscale image After scaling to the same size and overlaying, a three-channel visual taste fingerprint is obtained.

[0088] In this embodiment, each channel of the three-channel grayscale image is essentially transformed from a two-dimensional numerical matrix, where each element corresponds to a pixel in the image or a pixel grayscale value after scaling and interpolation. The specific execution process is as follows: For any matrix to be mapped, denoted as... , its first line, number Column elements are First, the matrix elements are mapped to using a min-max normalization method. Interval:

[0089] in, These are the normalized matrix elements. For matrix The minimum value in, For matrix The maximum value in, To prevent extremely small positive numbers with a denominator of zero, the normalized values ​​are then converted to grayscale values.

[0090] in, For the image at position The grayscale value at each location ranges from 0 to 255. A larger grayscale value indicates a higher feature strength, relationship strength, or recursive similarity corresponding to that location in the matrix; a smaller grayscale value indicates a lower corresponding value.

[0091] For the first channel grayscale image, its source is the event feature matrix. Event Feature Matrix The rows represent the sequence numbers of the response sub-events, and the columns represent the feature dimensions, such as morphological features, spectral features, cepstral features, time-frequency energy features, graph structure features, and order pattern complexity features. Since different feature dimensions have different dimensions and numerical ranges, the event feature matrix is ​​first normalized column-wise or as a whole to obtain a normalized event feature matrix. Then Each element in the image is mapped to a grayscale value between 0 and 255, resulting in the first channel grayscale image:

[0092] This channel is used to characterize the local taste response texture of each response sub-event across different feature dimensions.

[0093] The second channel grayscale image is derived from the event relationship matrix. Event Relationship Matrix The rows and columns represent the sequence numbers of the response sub-events, and the matrix elements... Indicates the first The response sub-event and the first The comprehensive relationship strength between response sub-events is determined by amplitude relationships, area relationships, cepstral relationships, and graphical relationships. (This refers to the event relationship matrix.) After normalization, the normalized result Convert to grayscale to obtain the second channel grayscale image:

[0094] This channel is used to characterize the global coupling and similarity relationships between different response sub-events.

[0095] The third channel grayscale image originates from a recursive texture matrix. Recursive texture matrix The rows and columns also represent the sequence numbers of the response sub-events, and the matrix elements... Indicates the first The first response sub-event and the first The recursive similarity of each response sub-event at the event centroid vector level. Because It is usually calculated using an exponential kernel function, and its value is generally between 0 and 1, so it can be directly used. The normalized similarity value can be used as the basis for further image contrast enhancement using minimum-maximum normalization. The third channel grayscale image is then obtained using the following formula:

[0096] This channel is used to characterize repetitive patterns, similar rhythms, and recursive structures across events in the entire taste response.

[0097] After obtaining the first channel grayscale image, the second channel grayscale image, and the third channel grayscale image, scale them all to the same size, for example, scale them all to a uniform size. Pixels. Scaling methods can include nearest neighbor interpolation, bilinear interpolation, or other conventional image scaling methods. Finally, the three channels are superimposed along their channel dimensions to form a three-channel visual taste fingerprint:

[0098] Therefore, the gray values ​​of the three grayscale channels are the image representation values ​​obtained after normalization and quantization of the corresponding matrix elements. Among them, the gray value of the first channel represents the intensity of the normalized multi-domain features, the gray value of the second channel represents the intensity of the normalized event relationships, and the gray value of the third channel represents the recursive similarity of events after normalization.

[0099] Taste fingerprints simultaneously preserve local texture, global relationships, and repetitive patterns across events, exhibiting excellent visual recognizability and category specificity. The first channel highlights local taste response texture, the second channel emphasizes the coupling strength between events, and the third channel highlights repetitive response patterns, enabling the generated fingerprints to be used for both human observation and subsequent machine recognition and database management.

[0100] The following is a specific implementation case of constructing a one-dimensional temporal response taste fingerprint based on a voltage-ampere type taste system: A voltammetric taste system was used to detect the test samples. The detection device included a working electrode, a reference electrode, a counter electrode, a constant potential control unit, a signal acquisition unit, and a temperature recording unit. The food sample was placed in the detection cell, and a potential sequence was applied according to a preset excitation program to obtain the current response signal of the working electrode. The combined taste signal was acquired at a fixed sampling frequency to obtain a one-dimensional current response sequence of length N. .

[0101] Due to the presence of high-frequency noise, baseline drift, and ambient temperature disturbances in the original current response sequence, a Savitzky-Golay filter with a window length of 9 and a fitting order of 3 was first used for smoothing. Then, an asymmetric least squares method was used to fit the baseline. Next, temperature drift correction was applied to the signal based on the recorded temperature values. Finally, robust normalization was performed using the median and absolute median difference to obtain the standardized sequence. After the above processing, the overall drift of the sequence is suppressed, and local transients are preserved, laying the foundation for subsequent event segmentation and deep feature extraction. The above preprocessing takes into account the actual complexity of food sample testing, and is especially suitable for sample systems with large differences in matrix conditions, such as wines, beverages, fermentation broths, flavoring liquids, and plant extracts.

[0102] For standardized sequences Calculate the first-order and second-order differences to construct the fused energy sequence. .exist Local maxima are sought within the sequence, and valid boundary points are selected based on thresholds and minimum time intervals. For the responses of the taste system under impulse, step, and combined excitations, boundary points typically correspond to excitation switching, polarization abrupt changes, and recovery processes. Based on the detected boundary points, the entire sequence is decomposed into K response sub-events, each corresponding to a micro-action unit in the taste response process. Each response sub-event is then uniformly resampled into 128 points to ensure a comparable, uniform length.

[0103] Figure 5 This is an image showing the results of uniform-length resampling and segment alignment based on actual collected samples. Figure 5 As can be seen, after event boundary detection, each response sub-event is mapped to an equal-length sequence, forming a resampled event matrix as shown below. Figure 6 As shown, this provides a unified input for subsequent multi-domain feature extraction. Figure 5 This is a superimposed graph of curves for each response sub-event after uniform-length resampling. The horizontal axis represents the resampling point number, typically from 1 to 128, indicating the relative time position of each response sub-event after uniform-length resampling. The vertical axis represents the normalized response amplitude, indicating the standardized response strength of the event at the corresponding resampling point. Curves of different colors represent different response sub-events obtained by event boundary detection; different colors are only used to distinguish event numbers. That is, each color corresponds to a response sub-event segmented from the original whole time-series signal by event boundary detection. Each event is resampled into 128 points, so they can be superimposed and compared on the same coordinate axis. Figure 5 As can be seen, response segments with different original durations are resampled and mapped onto a normalized time axis of the same length, thus allowing their morphological differences to be compared under a unified coordinate system. Figure 6 This is a heatmap of the resampled event matrix. The horizontal axis represents the resampled point index, and the vertical axis represents the event index. Each horizontal bar corresponds to an equal-length sequence of response sub-events. The color within a bar represents the normalized amplitude of that event at the corresponding resampled point, with color intensity corresponding to the magnitude of the normalized amplitude. Specifically, the color change from left to right within a bar represents the dynamic change in normalization within that event from start to end. The color differences between different bars represent the morphological and amplitude differences between different response sub-events. If K response sub-events are detected, and each sub-event is uniformly resampled to M points, then... Figure 6 This corresponds to a K×M event matrix. This matrix provides a unified input for subsequent modeling of morphological features, spectral features, cepstral features, graph structure features, and event relationships.

[0104] For each response sub-event, morphological features are first extracted, including positive and negative peak values, rise time constant, recovery time constant, signed area, and curvature energy, to characterize the significant morphology of the transient response. Secondly, short-time spectral features are extracted, including spectral centroid, spectral bandwidth, spectral roll-off point, spectral entropy, and spectral flux, to characterize the frequency distribution and energy transfer in the response. Then, electrochemical cepstral coefficients (ECC) are extracted; this is a key innovation of this invention. By performing a discrete cosine transform on the energy of the logarithmic filter bank, the hidden periodicity, slow recovery trend, and weak oscillation modes in the complex transient are compressed into low-dimensional cepstral coefficients. Next, three-layer wavelet packet energy features are extracted to characterize the energy distribution at different time scales. Then, graph structure features are constructed using the visualization method, extracting mean degree, clustering coefficient, graph entropy, and spectral radius, to characterize the global topological organization of the response morphology. Finally, permutation entropy, transition entropy, and statistical complexity are extracted to characterize the orderliness and complexity of the microscopic response sequence.

[0105] Figure 7This is a graph showing the ECC extraction results based on the electrochemical cepstral features of the actual collected samples. Figure 7 Figures (a), (b), (c), and (d) show the representative response sub-events, power spectrum, frequency-weight distribution of the non-uniform triangular filter bank, and bar chart of the electrochemical cepstral coefficients, respectively. These figures can be used to visually illustrate the mapping process of the one-dimensional temporal response of the taste system to the electrochemical cepstral feature space.

[0106] Figure 8 The resulting graph is constructed based on the graph structure features of actual collected samples. This graph shows representative response sub-events, natural visibility, adjacency matrix, and corresponding graph structure parameters, illustrating how the present invention converts the same representative one-dimensional taste response into a graph topology and extracts structural features. Figure 8 The middle figure (a) shows a representative response sub-event, with the horizontal axis representing the node number or resampling point number and the vertical axis representing the normalized response amplitude. Each sampling point on the curve is used as a node in the graph structure. Figure 8 Figure (b) is a natural visibility diagram. The meanings of its horizontal and vertical axes are the same as those in Figure (a). The lines connecting the nodes in the figure represent that the corresponding two sampling points satisfy the natural visibility diagram criteria and establish an edge connection relationship. That is, if the line connecting two sampling points is not occluded by the intermediate sampling point, then an edge is established between the two points. This line is used to characterize the topological visibility relationship between different peaks and valleys in the response curve. Dense lines indicate that the corresponding curve region has strong structural connectivity. Sparse lines indicate that local peaks, valleys or turns occlude the visibility relationship of the nodes. Figure 8 Figure (c) shows the adjacency matrix, where the x and y axes are node indices. Elements in the matrix represent whether a visual connection exists between any two nodes; a connection is denoted as 1, and no connection is denoted as 0. Since this visual graph is undirected, the adjacency matrix has a symmetric structure. Figure 8 The graph structure parameters calculated based on the adjacency matrix include: average degree, clustering coefficient, spectral radius, graph entropy, average path length, and degree distribution skewness. These parameters are used to intuitively and quantitatively characterize the overall topological organization and local structural complexity of the taste response curve. Specifically, average degree reflects the average number of connections between nodes; the clustering coefficient reflects the degree of local node clustering; the spectral radius reflects the overall connection strength of the graph; the graph entropy reflects the complexity of the connection distribution; the average path length reflects the overall reachability between nodes; and the degree distribution skewness reflects whether the connections are concentrated on a few key nodes.

[0107] All response sub-events are arranged chronologically to form an event feature matrix F. This matrix not only reflects the attributes of each local response fragment but also preserves macroscopic rhythmic information such as the initial strength followed by weakness, the initial depression followed by a rise, and the slowing recovery throughout the entire taste sequence. Furthermore, this invention does not directly use F as the final fingerprint but instead constructs an event relationship matrix R. The event relationship matrix can express the relative amplitude, relative area, relative cepstral angle, and relative spectral distance between events, essentially constructing a taste response grammar structure. Food samples of the same type not only have similar local fragment morphologies but also typically exhibit strong consistency in their cross-fragment relationships; therefore, introducing R can significantly improve the stability and specificity of the fingerprint.

[0108] After obtaining F and R, a three-channel taste fingerprint is constructed. First channel layer. Local feature textures used to represent event sequences; second channel layer Used to represent the strength and coupling patterns between events; third channel layer This is used to represent the recursive texture structure between event centroids. After scaling the three channels to a uniform size and then overlaying them, a visual taste fingerprint is obtained. I The texture distribution, diagonal structure, light-dark transitions, and local block patterns of different types of samples on this image show significant differences, thus it can serve as a taste-specific identification map. Different flavor profiles can exhibit different texture densities, diagonal aggregation degrees, block brightness distributions, and recursive repeating structures in the comprehensive taste quality three-channel fingerprint image, thus forming a taste map expression with discernible significance.

[0109] Figure 9 This is a schematic diagram of a three-channel taste fingerprint constructed based on actual collected samples. In the diagram, C1 shows the event feature matrix texture, C2 shows the event relationship matrix, and C3 shows the recursive texture matrix. The fused three-channel taste fingerprint is presented, which can be used to illustrate the collaborative coding effect of the present invention on the integrated local features of taste, global relationships and cross-event repetition patterns.

[0110] In one embodiment, step S8 is further included: to facilitate database storage and fast indexing, the present invention further visualizes the fingerprint map. Generate binary fingerprint code H The specific method involves performing a two-dimensional discrete cosine transform on the fused grayscale image, extracting low-frequency information, and then quantizing it according to the median threshold to generate a 255-bit binary sequence.

[0111] In this embodiment, the steps for generating the binary taste fingerprint code include: S81: Perform grayscale fusion on the three-channel visualized taste fingerprint to obtain a fused grayscale image; S82: Perform a two-dimensional discrete cosine transform on the fused grayscale image, extract the low-frequency coefficient value region, remove the DC component from the low-frequency coefficient value region, and then calculate the median threshold; wherein, the low-frequency coefficient value region is the region of coefficient values ​​with frequency values ​​lower than the set threshold. S83: Binarize the corresponding two-dimensional discrete cosine transform results of each coordinate position of the fused grayscale image according to the median threshold to obtain the binary values ​​of each coordinate position, forming a binary sequence.

[0112] In specific implementation, such as Figure 10 As shown, the three-channel taste fingerprint image is first fused with grayscale to obtain:

[0113] In the formula, To merge grayscale images, , , This is for layer weight. Then... Perform a two-dimensional discrete cosine transform, extract the low-frequency coefficient block from the upper left corner, remove the DC component, and then calculate the median threshold. And perform binary quantization:

[0114] in, The horizontal and vertical coordinate positions of the fused grayscale image after two-dimensional discrete cosine transform. The coefficient values ​​at each location are calculated. All binary results are expanded in a predetermined order to obtain the binary taste fingerprint code. :

[0115] This binary sequence maintains overall stability even with slight sample fluctuations and the presence of local noise, making it suitable for batch retrieval, similarity matching, and anti-counterfeiting tracking. This binary fingerprint can be used for sample comparison, database retrieval, batch identification, and similarity calculation.

[0116] Figure 11 This is a schematic diagram of the binary fingerprint code H generated based on actual collected samples. Figures (a), (b), and (c) sequentially show the grayscale fusion image, the low-frequency DCT coefficient block, and the distribution of the binarized fingerprint code, illustrating the process of generating machine-readable taste codes from a visualized taste fingerprint image in this invention.

[0117] In one embodiment, a food taste sensory fingerprint database is established. N repeated tests are performed on food samples of the same category, and the results are obtained according to steps S1 to S7. N Individual sample-level taste fingerprints N >1. Perform dimension-by-dimensional aggregation on its event feature matrix to obtain the category-level prototype feature matrix. :

[0118] Element-wise aggregation of its event relationship matrix yields a category-level prototype relationship matrix. :

[0119] in, This refers to the number of repeated measurements for the same type of sample.

[0120] A category-level prototype code is constructed using a bit-by-bit majority voting method for binary fingerprint codes.

[0121] The prototype fingerprint thus constructed retains intra-class commonalities while suppressing individual noise, and can be used for consistency evaluation of different batches of products, as well as for similarity comparison between new samples and standard samples.

[0122] In one embodiment, taste fingerprints are constructed for samples with different taste categories, concentration gradients, processing techniques, or storage stages to establish a food taste sensory fingerprint database. This database is used for subsequent food taste classification, authenticity verification, process tracking, batch stability evaluation, and abnormal sample identification. By establishing category-level prototype fingerprints, standard taste templates for sour, sweet, bitter, or complex taste samples can be further developed for rapid classification and stability evaluation of unknown samples.

[0123] Based on the same inventive concept, embodiments of the present invention also provide a food taste recognition method based on a food taste sensory data image fingerprint construction method, comprising the following steps: Using the first aspect of the food taste sensory data image fingerprint construction method, a taste fingerprint of the food sample to be tested is constructed. The taste fingerprint of the food sample to be tested is compared with at least one taste fingerprint in the three-channel visualized taste fingerprint database, and the distance between the two is calculated. Based on the relationship between distance and a preset threshold, the taste category or quality status of the food sample to be tested is determined.

[0124] In one embodiment, the comparison method includes visual image similarity, event feature matrix distance, relation matrix distance, and binary code Hamming distance. Through multi-level similarity fusion, it can be used to identify the taste category, processing differences, batch shifts, and anomalous samples in food products.

[0125] As an example, during sample comparison, two binary taste fingerprint codes and The degree of difference can be measured using the Hamming distance:

[0126] In the formula, The length of the binary fingerprint. and The two samples were respectively in the first... The coded value at each bit. The smaller the Hamming distance, the higher the similarity of the taste fingerprints between the two samples.

[0127] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0128] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing image fingerprints of food taste sensory data, characterized in that, Includes the following steps: S1: Collect one-dimensional time-series response data of food samples under the taste detection system; S2: Preprocess the one-dimensional time-series response data to obtain a standardized response sequence; S3: Based on the local mutation intensity of the standardized response sequence, perform event boundary detection, divide the entire response sequence into multiple response sub-events based on the event boundary points, and perform uniform length resampling on each response sub-event to obtain standardized response sub-events; S4: Extract multi-domain features for each standardized response sub-event, including at least morphological features, spectral features, cepstral features, time-frequency energy features, graph structure features, and order pattern complexity features, to form an event feature matrix; S5: Construct an event relationship matrix based on the amplitude relationship, area relationship, cepstral relationship, and spectral relationship between different events in the event feature matrix; S6: Construct a recursive texture matrix based on the centroid vectors of each standardized response sub-event; S7: Generate a three-channel visual taste fingerprint based on the event feature matrix, the event relationship matrix, and the recursive texture matrix.

2. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, The one-dimensional time-series response data are current response sequences, potential response sequences, or conductivity response sequences collected during the steady-state stage of the reference liquid, the sample stimulation stage, and the cleaning and regeneration detection stage.

3. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, The event boundary detection steps in step S3 include: calculating the first-order and second-order differences of the sequence amplitudes corresponding to the standardized response sub-events; fusing the first-order and second-order differences according to preset weights to obtain a fusion energy function; selecting local peak points in the fusion energy function that exceed a predetermined multiple of the mean standard deviation as candidate boundary points; and applying a minimum time interval constraint to obtain an effective boundary set.

4. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, In the event feature matrix of step S4: The morphological features include at least positive peak value, negative peak value, peak spacing, rise time constant, recovery time constant, signed area, curvature energy, and zero crossover delay; The spectral features include at least the spectral centroid, spectral bandwidth, spectral roll-off point, spectral flatness, spectral entropy, spectral flux, dominant modulation frequency, and high-to-low frequency energy ratio. The cepstral features are obtained by electrochemical cepstral coefficients, which are obtained by integrating the amplitude spectrum of the response sub-events using a non-uniform filter bank, logarithmic compression, and discrete cosine transform. The time-frequency energy features are decomposed using a three-layer wavelet packet decomposition. The energy of each sub-band node is normalized to obtain the wavelet packet energy distribution vector. The graph structure features are obtained by converting one-dimensional response sub-events into an undirected graph using the visual graph method, and at least one of the following is extracted as graph structure features: average degree, clustering coefficient, graph radius, graph entropy, degree distribution skewness, and average path length. The order pattern complexity features include at least permutation entropy, order transition entropy, and statistical complexity.

5. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, The elements of the event relationship matrix in step S5 are obtained by weighting the amplitude ratio, area ratio, cepstral angle, and spectral distance between any two standardized response sub-events according to preset weights.

6. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, The steps in constructing the recursive texture matrix in step S6 include: S61: For each standardized response sub-event, calculate the normalized time centroid and amplitude centroid to form the event centroid vector; S62: Calculate the distance between the centroid vectors of any two events, and map the distance to a similarity to obtain a recursive texture matrix.

7. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, Step S7 includes: The event feature matrix is ​​arranged in the order of event number and event feature dimension. After normalization and grayscale quantization of the event features, it is mapped to a first channel grayscale image. The event relationship matrix is ​​arranged in the order of event number and relationship strength dimension. The relationship strength is normalized and grayscaled, and then mapped to a second channel grayscale image. The recursive texture matrix is ​​arranged according to the event sequence number and the recursive similarity dimension. The recursive similarity is normalized and grayscale is then mapped to a third-channel grayscale image. The first channel grayscale image, the second channel grayscale image, and the third channel grayscale image are scaled to the same size and then superimposed to obtain a three-channel visual taste fingerprint image.

8. The method for constructing image fingerprints of food taste sensory data as described in claim 1, characterized in that, It also includes step S8: quantizing the three-channel visualized taste fingerprint to generate a binary taste fingerprint code.

9. The method for constructing image fingerprints of food taste sensory data as described in claim 8, characterized in that, The steps for generating the binary taste fingerprint include: S81: Perform grayscale fusion on the three-channel visualized taste fingerprint to obtain a fused grayscale image; S82: Perform a two-dimensional discrete cosine transform on the fused grayscale image, extract the low-frequency coefficient value region, remove the DC component from the low-frequency coefficient value region, and then calculate the median threshold; wherein, the low-frequency coefficient value region is the coefficient value region with a frequency value lower than the set threshold. S83: Binarize the corresponding two-dimensional discrete cosine transform results of each coordinate position of the fused grayscale image according to the median threshold to obtain the binary values ​​of each coordinate position, forming a binary sequence.

10. A method for food taste recognition, characterized in that, A three-channel visualized taste fingerprint database for different foods is constructed using the food taste sensory data image fingerprint construction method according to any one of claims 1 to 9; comprising the following steps: Using the food taste sensory data image fingerprint construction method according to any one of claims 1 to 9, a taste fingerprint of the food sample to be tested is constructed. The taste fingerprint of the food sample to be tested is compared with at least one taste fingerprint in the three-channel visualized taste fingerprint database, and the distance between the two is calculated. Based on the relationship between the distance and the preset threshold, the taste category or quality status of the food sample to be tested is determined.

Images