Food sensory evaluation method based on attention mechanism

Abstract

The invention relates to a food sensory evaluation method and device based on an attention mechanism, equipment and a medium. The method comprises the following steps: acquiring sensory information such as vision, smell, taste and touch of a food sample through multi-modal data acquisition equipment, and performing standardization processing to obtain a sensory feature matrix; generating a sensory feature vector set by using a pre-established multi-modal feature extraction model; modeling each sensory dimension relationship by means of an attention layer in a deep learning network, and calculating to obtain an attention weight of each sensory feature; and performing weighted combination on the sensory feature vector set based on the weight to generate a comprehensive sensory feature basic representation. According to the method, the association among the multi-modal sensory characteristics is effectively modeled through an attention mechanism, and a scientific and efficient solution is provided for food sensory evaluation.

Description

Technical Field

[0001] The present invention belongs to the field of food sensory evaluation, and in particular relates to a food sensory evaluation method based on an attention mechanism. Background Art

[0002] With the development of food sensory evaluation technology, multimodal data processing, and deep learning technology, automated sensory evaluation technology based on computer technology has emerged. This technology attempts to achieve quantitative analysis of food sensory characteristics by integrating multimodal data such as vision and smell, which in turn leads to the current traditional food sensory evaluation method or analysis method based on single-modal data. In traditional technologies, it is often necessary to rely on manual sensory evaluation teams to score based on subjective experience, or use a single sensor to independently detect the sensory characteristics of a certain dimension of food (such as visual color, olfactory concentration), and then generate evaluation results through a simple weighting method. However, the current traditional evaluation methods or single-modal analysis methods have the problems of strong subjectivity and poor consistency in manual evaluation, and are unable to effectively model the correlation between multiple sensory dimensions (such as vision and taste). As a result, the evaluation of the comprehensive sensory characteristics of food is insufficiently accurate, making it difficult to meet the needs of intelligent and precise food quality assessment. Summary of the Invention

[0003] Based on this, it is necessary to provide a food sensory evaluation method, device, equipment and medium based on an attention mechanism that can solve the above problems.

[0004] In a first aspect, the present application provides a food sensory evaluation method based on an attention mechanism, comprising:

[0005] Sensory information is acquired from food samples using multimodal data acquisition equipment and standardized to obtain a sensory feature matrix; sensory information includes visual, olfactory, taste, and tactile information;

[0006] For the sensory feature matrix, a pre-established multimodal feature extraction model is used to process and generate a sensory feature vector set;

[0007] The attention layer in the pre-built deep learning network is used to model the relationship between the sensory dimensions in the sensory feature vector set and calculate the attention weight of each sensory feature;

[0008] Based on the attention weights, the sensory feature vector sets are weightedly combined to generate a comprehensive sensory feature base representation.

[0009] In one embodiment, the method further comprises:

[0010] Detecting the element value of each sensory feature vector in the comprehensive sensory feature base representation;

[0011] When the element value of any sensory feature vector exceeds the preset threshold value of the sensory dimension corresponding to the element value, the element value of the sensory feature vector is corrected using a preset correction strategy;

[0012] The element values ​​of the corrected sensory feature vector are used to update the basic representation of the comprehensive sensory feature to obtain the verified comprehensive sensory feature representation.

[0013] In one embodiment, the method further comprises:

[0014] The verified comprehensive sensory characteristics are evaluated in multiple dimensions; the multidimensional evaluation includes visual pleasure evaluation, olfactory appeal evaluation, taste coordination evaluation, and tactile satisfaction evaluation.

[0015] The visual pleasure score and the olfactory appeal score are combined to generate the product aesthetics index, and the taste harmony score and the tactile satisfaction score are combined to generate the taste harmony index.

[0016] Based on the product aesthetic index and taste harmony index, the basic comprehensive sensory score is output through the pre-established sensory decision matrix.

[0017] In one embodiment, the method further comprises:

[0018] Obtain a dataset of consumers’ experience ratings of the same food samples;

[0019] By comparing the deviation distribution of basic comprehensive sensory rating and experience rating datasets, we can identify the deviation in attention weight allocation.

[0020] The attention weight of each sensory feature is adjusted based on the attention weight distribution bias.

[0021] In one embodiment, the method further comprises:

[0022] Establish a mapping relationship between the basic comprehensive sensory score and the experience score dataset;

[0023] Based on the mapping relationship, the basic comprehensive sensory score representation is synchronously transformed and processed to output the food sensory score;

[0024] Combined with food sensory scores and mapping relationships, key feature contribution analysis is performed to generate a key feature contribution analysis report.

[0025] In one embodiment, the key feature contribution analysis is implemented by the following mathematical relationship:

[0026]

[0027] Where n is the total number of sensory features, j is the normalization coefficient index, η iis the key feature contribution of the i-th sensory feature, w i , is the adjusted attention weight of the i-th sensory feature, δ i =|s i , -μ| is the score offset of the i-th sensory feature under the mapping relationship, s i , is the sensory characteristic score of item i after synchronous transformation, is the average score of all sensory features, and k is the global average index.

[0028] In one embodiment, the method further comprises:

[0029] Generate a rating description rule set based on the feature contribution distribution in the key feature contribution analysis report;

[0030] Determine the rating label of the food sample based on the comparison result of the food sensory score with the preset grade threshold;

[0031] By integrating level labels and rating description rule sets, a food rating report text containing sensory feature analysis and optimization suggestions is generated through preset semantic templates.

[0032] In a second aspect, the present application also provides a food sensory evaluation device based on an attention mechanism, comprising:

[0033] An information acquisition and processing module is used to acquire sensory information from food samples through a multimodal data acquisition device and perform standardization processing to obtain a sensory feature matrix; sensory information includes visual, olfactory, taste, and tactile information;

[0034] A multimodal feature extraction module is used to process the sensory feature matrix using a pre-established multimodal feature extraction model to generate a sensory feature vector set;

[0035] An attention weight calculation module is used to model the relationship between the sensory dimensions in the sensory feature vector set using the attention layer in the pre-established deep learning network, and calculate the attention weight of each sensory feature;

[0036] The comprehensive feature generation module is used to perform weighted combination of the sensory feature vector set based on the attention weight to generate a basic representation of the comprehensive sensory feature.

[0037] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-mentioned food sensory evaluation method based on the attention mechanism.

[0038] In a fourth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned food sensory evaluation method based on the attention mechanism are implemented.

[0039] The above-mentioned food sensory evaluation method, device, computer equipment and storage medium based on the attention mechanism synchronously obtains the visual, olfactory, taste and tactile information of food samples through multimodal data acquisition equipment and standardizes it into a feature matrix. It generates a sensory feature vector set with the help of a pre-established multimodal feature extraction model, uses the attention layer in the deep learning network to model the relationship between each sensory dimension and calculate the attention weight, and generates a basic representation of comprehensive sensory features through weighted combination. This solution effectively avoids the limitations of traditional single-modal detection through the integrated collection and standardized processing of multimodal data, uses the attention mechanism to achieve quantitative modeling of the correlation between multiple sensory dimensions, solves the problems of strong subjectivity and poor consistency in manual evaluation, and improves the accuracy and objectivity of the comprehensive sensory feature evaluation of food through automated feature extraction and weighted combination strategies, meeting the needs of intelligent and precise food quality assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 This is a flow chart of a food sensory evaluation method based on an attention mechanism of the present invention;

[0042] Figure 2 This is a structural diagram of a food sensory evaluation device based on the attention mechanism of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0044] In one embodiment, Figure 1As shown, a food sensory evaluation method based on an attention mechanism is provided. This embodiment is explained by taking the application of this method to a terminal equipped with a multimodal data acquisition device as an example, wherein the terminal may include hardware carriers such as an industrial control terminal and an intelligent detection device. The multimodal data acquisition device integrates a visual image acquisition module, a gas sensor array, a taste sensor group and a tactile pressure sensor, and can be deployed in scenarios such as food production lines and quality inspection laboratories to obtain sensory information of food samples in real time. It should be understood that this method can also be applied to cloud servers or edge computing servers, and the multimodal feature extraction model and deep learning network operations are executed through the high-performance computing resources of the server; it can also be applied to a distributed system composed of a terminal and a server, and the terminal completes on-site data collection and preprocessing, and transmits the data to the server through the network for feature vector generation, attention weight calculation and comprehensive feature generation, so as to realize automated sensory evaluation in application scenarios such as online detection of food production processes, finished product quality assessment, and sensory optimization of new product development. In this embodiment, the method includes the following steps:

[0045] S01, obtain sensory information from food samples through multimodal data acquisition equipment, and perform standardization processing to obtain a sensory feature matrix; sensory information includes visual, olfactory, taste and tactile information.

[0046] Among them, through the detection device integrating multi-dimensional sensing functions, the visual, olfactory, taste, tactile and other multimodal sensory information of food samples are synchronously collected, and the collected raw data is subjected to standardized pre-processing, including but not limited to data cleaning, normalization, feature dimension alignment and other operations, to generate a structured sensory feature matrix, providing basic data input in a unified format for subsequent multimodal feature fusion. Among them, the multimodal data acquisition equipment includes various sensor combinations that can realize visual image acquisition, gas molecule recognition, taste signal perception and mechanical property detection. Through standardization processing, the dimensional differences and format deviations of different modal data are eliminated, and a feature representation space with cross-dimensional comparability is constructed.

[0047] S02, for the sensory feature matrix, a pre-established multimodal feature extraction model is used to process and generate a sensory feature vector set.

[0048] Among them, a pre-built and trained multimodal feature extraction model (optimized through large-scale multimodal food sensory data training and has the ability to jointly represent cross-modal features) is used to perform cross-dimensional feature extraction operations on the sensory feature matrix obtained after standardization. The multimodal feature extraction model includes but is not limited to deep learning models based on convolutional neural networks, recurrent neural networks, or autoencoder architectures. Through multi-level feature transformation and cross-modal fusion operations, it extracts discriminative abstract features from multi-dimensional feature matrices such as vision, smell, taste, and touch, and generates a structured sensory feature vector set. This vector set represents the key feature information of each sensory modality in the form of a low-dimensional dense vector, providing a standardized feature representation for subsequent multimodal association modeling based on the attention mechanism.

[0049] S03, using the attention layer in the pre-established deep learning network to model the relationship between the sensory dimensions in the sensory feature vector set, and calculate the attention weight of each sensory feature.

[0050] Among them, the attention layer structure in the pre-built and trained deep learning network (including the implementation of the attention mechanism based on architectures such as Transformer, convolutional neural network or recurrent neural network, whose parameters have been optimized through multimodal food sensory data training) is used to quantitatively model the feature correlation relationship of each sensory dimension (vision, smell, taste, touch) in the sensory feature vector set. Its attention layer calculates the correlation weight coefficient between each sensory feature vector by designing a cross-modal interaction mechanism (including but not limited to self-attention, cross-attention or hybrid attention architecture) to form a set of attention weights that can characterize the difference in feature importance. Through matrix operations, operations such as similarity calculation and weight normalization between feature vectors are realized, so that the generated attention weights can adaptively characterize the difference in contribution of different sensory features in the comprehensive evaluation, providing a quantitative basis for the subsequent weighted fusion of multimodal features.

[0051] S04, based on the attention weight, performs weighted combination on the sensory feature vector set to generate a basic representation of the comprehensive sensory features.

[0052] Among them, based on the attention weights, a weighted combination operation is performed on the sensory feature vector set (including operations such as linear combination or nonlinear transformation based on the weight matrix, retaining the key information of each sensory dimension and quantifying the correlation between the representation modalities), to achieve collaborative representation of multi-dimensional sensory information: by performing a dot product operation or matrix multiplication operation on each sensory feature vector and the corresponding attention weight, a fusion feature space containing weight distribution of multimodal information such as vision, smell, taste, and touch is formed, generating a comprehensive sensory feature basic representation that can comprehensively reflect the multi-sensory dimension correlation characteristics of food samples. This basic representation adaptively adjusts the contribution of each sensory feature through the attention mechanism, achieving a structured representation of the comprehensive sensory characteristics of food samples, and providing a unified feature expression space for subsequent sensory feature verification, multidimensional scoring, and comprehensive evaluation.

[0053] The above-mentioned food sensory evaluation method based on the attention mechanism synchronously obtains the visual, olfactory, taste and tactile information of food samples through multimodal data acquisition equipment and standardizes it into a feature matrix, uses a pre-established multimodal feature extraction model to generate a sensory feature vector set, uses the attention layer in the deep learning network to model the relationship between each sensory dimension and calculate the attention weight, and generates a basic representation of comprehensive sensory features through weighted combination. Through the integrated collection and standardized processing of multimodal data, it avoids the limitations of traditional single-modal detection, uses the attention mechanism to realize quantitative modeling of the correlation between multiple sensory dimensions, solves the problems of strong subjectivity and poor consistency of manual evaluation, and improves the accuracy and objectivity of the comprehensive sensory feature evaluation of food through automated feature extraction and weighted combination strategies to meet the needs of intelligent and precise food quality assessment.

[0054] In one embodiment, the method further comprises:

[0055] S11, detecting the element value of each sensory feature vector in the comprehensive sensory feature basic representation;

[0056] S12, when the element value of any sensory feature vector exceeds a preset threshold value of the sensory dimension corresponding to the element value, correcting the element value of the sensory feature vector using a preset correction strategy;

[0057] S13, using the element values ​​of the corrected sensory feature vector, updating the comprehensive sensory feature basic representation to obtain a verified comprehensive sensory feature representation.

[0058] Specifically, the value range of each element of the feature vector can be checked through preset detection rules (such as element-by-element traversal and batch matrix operations) to eliminate extreme values ​​caused by sensor failure, data acquisition errors or model calculation anomalies. When the element value of any sensory feature vector exceeds the preset threshold of its corresponding sensory dimension (such as the brightness value of the visual feature exceeds the upper limit of human eye recognition, and the sweetness value of the taste feature exceeds the conventional range of the food industry), the element value of the sensory feature vector is corrected according to the preset correction strategy (which may include: truncation method, forcibly limiting the element value exceeding the threshold to the threshold boundary, such as setting it to 255 when the brightness value exceeds 255; normalization method, linear transformation of the element value based on the normal value range of the sensory dimension; interpolation method, replacing abnormal values ​​by interpolation of adjacent valid feature values.). Through matrix operations (such as element-level replacement and weighted update), the corrected feature vector is integrated into the original comprehensive feature space to generate a new feature matrix. Through the detection-correction-update process design, combined with the preset threshold and correction strategy, the validity verification of multimodal sensory features is achieved.

[0059] In one embodiment, the method further comprises:

[0060] S21, perform multidimensional scoring on the verified comprehensive sensory feature representation; the multidimensional scoring includes visual pleasure scoring, olfactory appeal scoring, taste coordination scoring, and tactile satisfaction scoring;

[0061] S22, combining the visual pleasure score with the olfactory appeal score to generate a product aesthetics index, and combining the taste harmony score with the tactile satisfaction score to generate a taste harmony index;

[0062] S23, based on the product aesthetic index and taste harmony index, outputs the basic comprehensive sensory score through the pre-established sensory decision matrix.

[0063] For example, multidimensional scoring may include: a visual pleasure score, which quantifies human visual preferences based on image aesthetic features (such as color saturation, shape symmetry, and gloss) through computer vision algorithms (such as CNN image aesthetic models); an olfactory appeal score, which uses gas sensor array data (such as volatile organic compound concentration distribution) combined with an olfactory psychology model to calculate odor comfort; a taste harmony score, which uses taste sensors (such as sour, sweet, bitter, salty, and fresh sensors) to detect the balance of flavor types and evaluate the rationality of the ratio of each taste component; and a tactile satisfaction score, which quantifies the physical experience of oral or hand contact based on texture data (such as hardness, elasticity, and viscosity) collected by pressure sensors. Each dimension score can be independently calculated using a pre-trained unimodal evaluation model, outputting a standardized score in the range of 0-100 to achieve cross-dimensional comparability. The product aesthetics index integrates visual and olfactory scores, focusing on the appearance and odor presentation of the food; the taste harmony index integrates taste and tactile scores, focusing on the taste-tactile synergy experience when eating. By categorizing and aggregating highly correlated sensory dimensions (e.g., vision and smell jointly influence first impressions, while taste and touch determine the eating experience), we reduce cross-modal noise interference and improve the index's interpretability. A pre-established sensory decision matrix, derived from historical evaluation data and consumer research training, characterizes the differential contributions of the aesthetic index and taste index to the overall score. Data-driven weight allocation avoids the subjectivity of manually setting weights. By defining a multidimensional scoring system, a classification fusion strategy, and a data-driven decision matrix, we achieve a comprehensive score from sensory features.

[0064] In one embodiment, the method further comprises:

[0065] S31, obtain the consumer experience rating dataset for the same food sample;

[0066] S32, identifying the attention weight allocation bias by comparing the deviation distribution of the basic comprehensive sensory score and the experience score dataset;

[0067] S33, adjusting the attention weight of each sensory feature based on the attention weight distribution bias.

[0068] Specifically, sensory ratings of the same food samples from target consumers can be collected through standardized questionnaires, focus group tests, or online review platforms. The rating dimensions are aligned to visual pleasantness, olfactory appeal, taste coordination, and tactile satisfaction. After cleaning, normalization, and noise reduction, a structured experience rating dataset is formed. The absolute deviation between the basic comprehensive sensory rating and the mean consumer rating is calculated, and the deviation contribution of each sensory dimension is further analyzed: the deviation between the scores of single dimensions such as vision and smell and the mean of the corresponding consumer dimension is calculated dimension by dimension to form a deviation distribution vector. Based on the deviation distribution vector, for dimensions with significant deviations, the weights can be adjusted inversely according to the deviation ratio (for example, if the visual dimension deviation is 15%, its attention weight is multiplied by a correction coefficient of 0.85. Through smoothing constraints (introducing regularization terms to avoid drastic fluctuations in weights) and iterative optimization (through multiple rounds of deviation identification and weight adjustment cycles until the deviation converges to a preset threshold (such as 5%)), the adjustment logic is embedded in the attention layer of the deep learning network. The weight parameters can be updated through the backpropagation algorithm, or the weight matrix can be directly corrected using heuristic rules. Through a data-driven dynamic adjustment strategy, instead of manually setting weights based on experience, the model's adaptability to different food categories and consumer groups is improved, more in line with market demand, and the ability to generalize across categories is enhanced.

[0069] In one embodiment, the method further comprises:

[0070] S41, establishing a mapping relationship between the basic comprehensive sensory score and the experience score dataset;

[0071] S42, performing synchronous transformation processing on the basic comprehensive sensory score representation based on the mapping relationship, and outputting the food sensory score;

[0072] S43, combining the food sensory scores and mapping relationships, performs key feature contribution analysis and generates a key feature contribution analysis report.

[0073] For example, supervised learning methods (such as neural networks, random forests, or linear regression) can be used, with a basic comprehensive sensory score as input and the mean of a consumer experience score dataset as output. A mapping function is trained to obtain a mapping relationship. The mapping relationship can include linear or nonlinear transformations (such as fitting complex nonlinear relationships through neural networks). Model parameters are optimized through cross-validation to ensure that the mapping error is ≤ a preset threshold (such as 3%). This bridges the scale difference between automated scoring and consumer subjective scoring (e.g., model output is 0-10 points, consumer rating is 0-100 points), and the basic comprehensive sensory score is converted into a scoring system that conforms to market perception. The basic comprehensive sensory score representation is synchronously transformed dimension by dimension through the mapping relationship to obtain a food sensory score that directly corresponds to the evaluation scale familiar to consumers. Based on the mapping relationship and food sensory score, a weight-bias joint analysis model can be used to quantify the contribution of each sensory feature. The feature contribution is calculated using the adjusted attention weight and the mapped score offset. A feature priority list is generated by sorting the contribution. The impact of the feature can be annotated with business semantics (e.g., low taste coordination leads to a 2.3-point decrease in the overall score), forming a structured report containing data indicators.

[0074] In one embodiment, S51, key feature contribution analysis is implemented by the following mathematical relationship:

[0075]

[0076] Where n is the total number of sensory features, j is the normalization coefficient index, η i is the key feature contribution of the i-th sensory feature, w i , is the adjusted attention weight of the i-th sensory feature, δ i =|s i , -μ| is the score offset of the i-th sensory feature under the mapping relationship, s i , is the sensory characteristic score of item i after synchronous transformation, is the average score of all sensory features, and k is the global average index.

[0077] Specifically, in this formula, the numerator w i ,·δ i represents the importance-bias joint effect of the i-th feature, and the adjusted attention weight w of the i-th sensory feature i , evaluate the score offset δ of the i-th sensory feature under the mapping relationship i The influence of w i , high and δ i When the numerator is large, the larger the value, the more critical the feature is to the overall score; the denominator The sum of the combined effects of all features is used to normalize the numerator to a percentage range. By summing the denominators, the contribution metrics are made comparable across the board. A mathematical formula combines the adjusted attention weights with the score offset. This utilizes a normalized calculation of a two-factor product, a dynamic deviation benchmark based on the average score, and a contribution output in percentage form to quantify the actual impact of each sensory characteristic on the overall evaluation, providing a data-driven decision-making basis for food quality optimization.

[0078] In one embodiment, the method further comprises:

[0079] S61, generating a rating description rule set based on the feature contribution distribution in the key feature contribution analysis report;

[0080] S62, determining a rating label for the food sample based on a comparison result of the food sensory score with a preset rating threshold;

[0081] S63, fusion-level label and rating description rule set, generates food rating report text containing sensory feature analysis and optimization suggestions through preset semantic templates.

[0082] For example, based on the contribution ranking of each feature in the key feature contribution analysis report (e.g., taste coordination contributes 35%, visual pleasure contributes 28%), combined with industry evaluation standards, a contribution-impact description mapping rule is established: high-contribution features (η>25%) are mapped to core impact descriptions (e.g., taste coordination is the main influencing factor, and its deviation causes the overall score to drop by 2.3 points); low-contribution features (η<10%) are mapped to secondary impact descriptions (e.g., tactile satisfaction has a minor impact and can be optimized later). This converts quantified contribution data into descriptive rules with business semantics, providing semantic material for report generation. Preset grade thresholds (e.g., 90-100 points for excellent, 80-89 points for good, etc.) can be set based on food category characteristics (e.g., the difference in scoring standards between baked goods and beverages) and market demand. The food sensory score is compared with the threshold range to determine the corresponding rating label (e.g., food sensory score = 85 → rating label good). Multi-dimensional ratings (e.g., comprehensive ratings and individual dimension ratings) can be performed simultaneously. For example, a composite label with a good comprehensive rating and an excellent visual dimension rating can be generated simultaneously to determine the rating label for a food sample. Based on a preset semantic template, the rating label is inserted into a fixed position in the template, and feature analysis and recommendation content are dynamically filled in according to the rule set to form a complete natural language text. Through a three-layer conversion mechanism of feature contribution, rating rules, and semantic templates, the automated generation of food sensory evaluation reports is achieved.

[0083] The above-mentioned food sensory evaluation method based on the attention mechanism synchronously obtains the visual, olfactory, taste and tactile information of food samples through multimodal data acquisition equipment and standardizes it into a feature matrix, uses a pre-established multimodal feature extraction model to generate a sensory feature vector set, uses the attention layer in the deep learning network to model the relationship between each sensory dimension and calculate the attention weight, and generates a basic representation of comprehensive sensory features through weighted combination. By detecting and correcting the feature vector element values, multi-dimensional score fusion is used to generate the product aesthetic index and taste harmony index, and the basic comprehensive sensory score is output in combination with the pre-established sensory decision matrix. The comparison deviation of the consumer experience score data set is obtained to adjust the attention weight, a score mapping relationship is established and the influence of each feature is quantified through the key feature contribution analysis formula, a rating rule set is generated based on the contribution distribution, the rating label is determined, and the semantic template is integrated to generate a rating report containing optimization suggestions. Through multimodal data integration and standardized processing, the limitations of traditional single-modal detection are avoided, the attention mechanism is used to quantitatively model the multi-sensory dimension association, and the weight distribution is dynamically optimized in combination with consumer feedback. The evaluation accuracy is improved through closed-loop correction and structured scoring system. With the help of mathematical models and semantic transformation, decision-making reports that can guide production are generated, which effectively solves the problems of strong subjectivity, poor consistency and insufficient multi-dimensional correlation modeling in manual evaluation, and meets the needs of intelligent and precise food quality assessment.

[0084] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0085] Based on the same inventive concept, the embodiments of the present application also provide a food sensory evaluation device based on an attention mechanism for implementing the aforementioned food sensory evaluation method based on an attention mechanism. The implementation solution provided by this device is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of a food sensory evaluation device based on an attention mechanism provided below can be found in the limitations of the food sensory evaluation method based on an attention mechanism above, and will not be repeated here.

[0086] In an exemplary embodiment, Figure 2As shown, a food sensory evaluation device based on an attention mechanism is provided, comprising:

[0087] The information acquisition and processing module 101 is used to acquire sensory information from food samples through a multimodal data acquisition device and perform normalization processing to obtain a sensory feature matrix; the sensory information includes visual, olfactory, taste and tactile information;

[0088] The multimodal feature extraction module 102 is used to process the sensory feature matrix using a pre-established multimodal feature extraction model to generate a sensory feature vector set;

[0089] An attention weight calculation module 103 is used to model the relationship between the sensory dimensions in the sensory feature vector set using the attention layer in the pre-established deep learning network, and calculate the attention weight of each sensory feature;

[0090] The comprehensive feature generation module 104 is used to perform weighted combination on the sensory feature vector set based on the attention weight to generate a basic representation of the comprehensive sensory feature.

[0091] In one embodiment, the comprehensive feature generation module 104 is further configured to:

[0092] Detecting the element value of each sensory feature vector in the comprehensive sensory feature base representation;

[0093] When the element value of any sensory feature vector exceeds the preset threshold value of the sensory dimension corresponding to the element value, the element value of the sensory feature vector is corrected using a preset correction strategy;

[0094] The element values ​​of the corrected sensory feature vector are used to update the basic representation of the comprehensive sensory feature to obtain the verified comprehensive sensory feature representation.

[0095] In one embodiment, the comprehensive feature generation module 104 is further configured to:

[0096] The verified comprehensive sensory characteristics are evaluated in multiple dimensions; the multidimensional evaluation includes visual pleasure evaluation, olfactory appeal evaluation, taste coordination evaluation, and tactile satisfaction evaluation.

[0097] The visual pleasure score and the olfactory appeal score are combined to generate the product aesthetics index, and the taste harmony score and the tactile satisfaction score are combined to generate the taste harmony index.

[0098] Based on the product aesthetic index and taste harmony index, the basic comprehensive sensory score is output through the pre-established sensory decision matrix.

[0099] In one embodiment, the attention weight calculation module 103 is further configured to:

[0100] Obtain a dataset of consumers’ experience ratings of the same food samples;

[0101] By comparing the deviation distribution of basic comprehensive sensory rating and experience rating datasets, we can identify the deviation in attention weight allocation.

[0102] The attention weight of each sensory feature is adjusted based on the attention weight distribution bias.

[0103] In one embodiment, a sensory evaluation module is further included for:

[0104] Establish a mapping relationship between the basic comprehensive sensory score and the experience score dataset;

[0105] Based on the mapping relationship, the basic comprehensive sensory score representation is synchronously transformed and processed to output the food sensory score;

[0106] Combined with food sensory scores and mapping relationships, key feature contribution analysis is performed to generate a key feature contribution analysis report.

[0107] In one embodiment, the key feature contribution analysis is implemented by the following mathematical relationship:

[0108]

[0109] Where n is the total number of sensory features, j is the normalization coefficient index, η i is the key feature contribution of the i-th sensory feature, w i , is the adjusted attention weight of the i-th sensory feature, δ i =|s i , -μ| is the score offset of the i-th sensory feature under the mapping relationship, s i , is the sensory characteristic score of item i after synchronous transformation, is the average score of all sensory features, and k is the global average index.

[0110] In one embodiment, the sensory evaluation module is further configured to:

[0111] Generate a rating description rule set based on the feature contribution distribution in the key feature contribution analysis report;

[0112] Determine the rating label of the food sample based on the comparison result of the food sensory score with the preset grade threshold;

[0113] By integrating level labels and rating description rule sets, a food rating report text containing sensory feature analysis and optimization suggestions is generated through preset semantic templates.

[0114] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the food sensory evaluation method based on the attention mechanism as described above are implemented.

[0115] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0116] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0117] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A food sensory evaluation method based on attention mechanism, characterized in that: The method comprises: Acquiring sensory information from food samples using a multimodal data acquisition device and performing standardization processing to obtain a sensory feature matrix; the sensory information includes visual, olfactory, taste, and tactile information; For the sensory feature matrix, a pre-established multimodal feature extraction model is used to process and generate a sensory feature vector set; Using the attention layer in a pre-established deep learning network to model the relationship between the sensory dimensions in the sensory feature vector set, and calculating the attention weight of each sensory feature; Based on the attention weights, the sensory feature vector sets are weightedly combined to generate a comprehensive sensory feature basic representation.

2. The method according to claim 1, characterized in that The method further comprises: detecting the element value of each sensory feature vector in the comprehensive sensory feature basic representation; When the element value of any of the sensory feature vectors exceeds a preset threshold value of the sensory dimension corresponding to the element value, a preset correction strategy is used to correct the element value of the sensory feature vector; The element values ​​of the corrected sensory feature vector are used to update the comprehensive sensory feature basic representation to obtain a verified comprehensive sensory feature representation.

3. The method according to claim 2, characterized in that The method further comprises: Performing a multi-dimensional scoring on the verified comprehensive sensory feature representation; the multi-dimensional scoring includes a visual pleasure score, an olfactory appeal score, a taste coordination score, and a tactile satisfaction score; The visual pleasure score and the olfactory appeal score are combined to generate a product aesthetics index, and the taste harmony score and the tactile satisfaction score are combined to generate a taste harmony index; Based on the product aesthetic index and taste harmony index, a basic comprehensive sensory score is output through a pre-established sensory decision matrix.

4. The method according to claim 3, characterized in that The method further comprises: Obtain a dataset of consumers’ experience ratings of the same food samples; Identifying attention weight allocation deviations by comparing the deviation distributions of the basic comprehensive sensory score and the experience score dataset; The attention weights of the sensory features are adjusted based on the attention weight allocation deviation.

5. The method according to claim 4, characterized in that The method further comprises: Establishing a mapping relationship between the basic comprehensive sensory score and the experience score dataset; Performing synchronous conversion processing on the basic comprehensive sensory score representation based on the mapping relationship, and outputting the food sensory score; Combine the food sensory scores and the mapping relationship to perform key feature contribution analysis and generate a key feature contribution analysis report.

6. The method according to claim 5, characterized in that The key feature contribution analysis is achieved through the following mathematical relationship: Where n is the total number of sensory features, j is the normalization coefficient index, η i is the key feature contribution of the i-th sensory feature, w i ， is the adjusted attention weight of the i-th sensory feature, δ i =|s i ， -μ| is the score offset of the i-th sensory feature under the mapping relationship, s i ， is the score of the i-th sensory characteristic after the synchronous transformation process, is the average score of all sensory features, and k is the global average index.

7. The method according to claim 5, characterized in that The method further comprises: generating a rating description rule set based on the feature contribution distribution in the key feature contribution analysis report; Determining a rating label for the food sample based on a comparison result of the food sensory score with a preset rating threshold; The level labels are integrated with the rating description rule set, and a food rating report text including sensory feature analysis and optimization suggestions is generated through a preset semantic template.

8. A food sensory evaluation device based on attention mechanism, characterized in that: The device comprises: An information acquisition and processing module is used to acquire sensory information from food samples through a multimodal data acquisition device and perform standardization processing to obtain a sensory feature matrix; the sensory information includes visual, olfactory, taste and tactile information; A multimodal feature extraction module, configured to process the sensory feature matrix using a pre-established multimodal feature extraction model to generate a sensory feature vector set; An attention weight calculation module, configured to use an attention layer in a pre-established deep learning network to model the relationship between the sensory dimensions in the sensory feature vector set and calculate the attention weight of each sensory feature; The comprehensive feature generation module is used to perform weighted combination on the sensory feature vector set based on the attention weight to generate a basic representation of comprehensive sensory features.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images