A method and system for non-destructive testing of meat

Abstract

The present application relates to the field of meat nondestructive testing, and particularly relates to a meat nondestructive testing method and system, comprising: obtaining a meat nondestructive testing model, the model is trained by the following steps: obtaining a hyperspectral image sample of sample meat, learning labels are set on the hyperspectral image sample; obtaining at least one auxiliary data such as drop loss rate, cooking loss rate, shear force, chewiness, etc.; performing food processing on multiple portions of sample meat to obtain meat products; recording the sensory scores given by a sensory evaluation personnel to the meat products; generating a training set based on the above data, and inputting the training set into a multi-modal fusion model for model training, and obtaining a corresponding meat nondestructive testing model; obtaining a to-be-tested hyperspectral image of to-be-tested meat, inputting the to-be-tested hyperspectral image into the meat nondestructive testing model, and then obtaining a predicted sensory score of the to-be-tested meat. The present application can directly output a predicted sensory score of meat based on a hyperspectral image, and further provides a professional comprehensive evaluation index for the mouthfeel of meat.

Description

Technical Field

[0001] This invention relates to the field of meat testing technology, specifically to a non-destructive testing method and system for meat. Background Technology

[0002] Currently, my country's frozen food industry does not have specific quality requirements for frozen foods. For example, GB31646-2018, the "National Food Safety Standard for Hygienic Specifications for the Production and Operation of Quick-Frozen Foods," only includes "Appendix A: Microbial Monitoring Requirements for the Processing of Quick-Frozen Cooked Foods." It only provides hygienic specifications for the production process of quick-frozen foods (including compliance with relevant provisions in GB14881-2013), without providing quality standards for frozen foods or offering reliable methods for judging the freshness of frozen foods.

[0003] However, with the gradual development of cold chain logistics, the safety and quality monitoring of frozen foods within this logistics chain has become crucial. Taking beef as an example, beef freezing processing includes direct freezing, where frozen beef is sold directly under frozen conditions after cold chain transportation, or quick-freezing frozen beef followed by cold chain transportation to its destination, where it is thawed and sold as fresh beef – a process called "frozen-to-fresh" conversion. However, due to freezing processes or storage time, thawed beef may suffer from significant nutrient loss and poor taste.

[0004] Currently, existing technologies attempt to assess meat freshness using hyperspectral imaging: For example, patent application CN202411775351.X discloses a deep learning-based multimodal fusion detection method for pork tenderloin freshness, which includes the following steps: preparing samples for a pork tenderloin freshness classification model; collecting HSI data, ultraviolet fluorescence image data, electronic nose data, and LF-NMR data of pork tenderloin during storage, and preprocessing the data; designing feature extraction modules for ultraviolet fluorescence image data and HSI data respectively; designing a data augmentation module based on variational autoencoder to enhance the ability to identify the boundaries of each category of sample freshness; designing an improved dung beetle optimization algorithm to enhance the model's hyperparameter optimization ability; introducing UV-CNN module, HSI-CNN module, AAVAE module, and EDBO optimization algorithm into the BiLSTM network structure to construct a CNN-AAVAE-EDBO-BiLSTM classification model for pork tenderloin freshness; and testing and evaluating the accuracy of the pork tenderloin classification model.

[0005] For example, patent application CN202011064908.0 discloses a method for rapidly characterizing the umami intensity of salmon based on hyperspectral imaging technology. First, salmon samples are prepared, and hyperspectral images are acquired using a hyperspectral imaging acquisition system. After acquisition, the umami intensity index of the samples is determined using the national standard method. Then, ENVI4.5 software is used to extract the spectral values ​​of the region of interest in the center of the sample, and the spectral values ​​are preprocessed. Next, different variable screening methods are used to extract feature variables from the preprocessed spectral values, and samples composed of the extracted feature variables are used as new samples. A sample set partitioning method is used to partition the new samples, resulting in a re-partitioned sample set. A salmon umami intensity index prediction model is established using the partitioned sample set, thereby realizing the visualization of the umami intensity index distribution.

[0006] However, the applicant noted that existing freshness prediction methods are insufficient for a precise assessment of the freshness or quality of meat, and therefore cannot meet the current testing application needs of cold chain logistics. Summary of the Invention

[0007] The purpose of this invention is to provide a non-destructive testing method that partially solves or alleviates the above-mentioned shortcomings in the prior art, and enables non-destructive testing of the freshness of meat.

[0008] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution: A first aspect of the present invention is to provide a non-destructive testing method for meat, comprising: S201, Obtain a non-destructive testing model for meat, wherein the non-destructive testing model is trained using the following steps: S2011, acquire multiple hyperspectral images of multiple meat samples to generate a hyperspectral image set, wherein the hyperspectral image samples are provided with a first label and a second label, the first label is used to mark volatile basic nitrogen information, and the second label is used to mark at least one auxiliary physicochemical index; S2012, Obtain an auxiliary sample, wherein the auxiliary sample includes at least one of the following auxiliary data of the sample meat: drip loss rate, cooking loss rate, shear force, and chewiness. S2013, The multiple meat samples are processed to obtain meat products; S2014, record at least one sensory rating given by the taster to the meat product; S2015, Generate a training set, the training set including multiple training sample groups, the training sample groups including: hyperspectral map samples, and the hyperspectral map samples are respectively associated with a type of auxiliary sample information and sensory scores; S2016, The training set is input into the multimodal fusion model for model training. After the training process converges, the corresponding non-destructive meat detection model is obtained. S202, acquire at least one hyperspectral image of the meat to be tested, input the at least one hyperspectral image of the meat to be tested into the non-destructive testing model of the meat, and output the corresponding predicted sensory score from the non-destructive testing model of the meat.

[0009] In some embodiments, S2016 includes the step: S20161, The hyperspectral image sample is obtained; S20162, at least one group of auxiliary physicochemical indicators is screened from the hyperspectral sample; S20163, at least one auxiliary data group is selected from the aforementioned auxiliary samples; S20164, Generate multiple condition combinations based on the volatile basic nitrogen information, at least one combination of auxiliary physicochemical indicators, and at least one combination of auxiliary data; S20165, The multimodal fusion model is used to learn the training set by combining multiple conditions to obtain multiple non-destructive detection models for meat.

[0010] In some embodiments, S20165 includes the step of: A first training subset and a second training subset are selected from the training set, and the first training subset is trained with the multimodal fusion model under at least two combination conditions, thereby obtaining multiple first models. Calculate the accuracy of at least two first predictions for the at least two first models; A second model that meets a first screening rule is selected from multiple first models, wherein the first screening rule requires that the first prediction accuracy of the second model is greater than a first accuracy threshold. The second model is optimized using the second training subset; Calculate the accuracy of at least two second predictions for the at least two second models; When the second prediction accuracy is higher than the second accuracy threshold, at least one of the corresponding second models will be output as the non-destructive testing model for meat.

[0011] In some embodiments, S202 includes: Multiple hyperspectral images to be tested are input into at least two non-destructive testing models for meat, and at least two sets of non-destructive testing results are obtained accordingly. The hyperspectral images to be tested are marked with category information, which includes: meat variety and / or food processing method. Calculate the third prediction accuracy for at least two sets of nondestructive testing results; When the difference between two of the third prediction accuracies is greater than the third accuracy threshold, the category information of the hyperspectral image with the higher third prediction accuracy is identified as the update information; wherein, the non-destructive meat detection model records at least one weight level, and the weight level is associated with category information; The updated information is used to update the recommended weights of the nondestructive testing model; The steps for updating the recommendation weights include: Obtain the set value of the weight level corresponding to the updated information; Increase the set value according to the set rules.

[0012] In some embodiments, at least one hyperspectral image to be measured includes: At least one first hyperspectral image covering a first area region of the meat to be tested, and at least one second hyperspectral image covering a second area region of the meat to be tested; wherein the first area region is larger than the second area region.

[0013] In some embodiments, S202 includes: Acquire at least one first hyperspectral image covering a first area region of the meat to be tested; Identify the spectral reflectance data of at least three sub-regions in the first hyperspectral image to be measured; Based on the spectral reflectance data, at least three target curves corresponding to the reflectance variation with wavelength are plotted respectively; Calculate the similarity index between any two of the at least three target curves; When the similarity index between one of the target curves and at least two other target curves is less than the first similarity threshold, the sub-region to be tested corresponding to the target curve is identified as the region of interest. A new second hyperspectral image to be measured is obtained for the region of interest; The first and second hyperspectral images to be tested are input together into the non-destructive testing model for meat.

[0014] In some embodiments, the sample meat includes: a first sample meat that has undergone freezing treatment, and a second sample meat that has not undergone freezing treatment.

[0015] In some embodiments, the nondestructive testing model is further used to output a frozen freshness value, which is the ratio of the sensory score of the sample meat before freezing to the sensory score after freezing.

[0016] The present invention also provides a non-destructive testing system for meat, comprising: The model module is used to obtain a non-destructive testing model for meat. This meat non-destructive testing model is created using a model creation module, which includes: The first sample acquisition unit is used to acquire multiple hyperspectral image samples of multiple meat samples to generate a hyperspectral image set. The hyperspectral image samples are provided with a first label and a second label. The first label is used to mark volatile basic nitrogen information, and the second label is used to mark at least one auxiliary physicochemical index. The second sample acquisition unit is used to acquire a type of auxiliary sample, which includes at least one type of auxiliary data of the sample meat, namely: drip loss rate, cooking loss rate, shear force, and chewiness. The scoring and recording unit is used to process the multiple meat samples to obtain meat products and record at least one sensory score given by the tasters to the meat products. A training set generation unit is used to generate a training set, which includes multiple training sample groups, each of which includes hyperspectral image samples and is associated with a type of auxiliary sample information and a sensory score. The training unit is used to input the training set into the multimodal fusion model for model training. After the training process converges, the corresponding non-destructive meat detection model is obtained. The output module is used to acquire at least one hyperspectral image of the meat to be tested, input the at least one hyperspectral image to the non-destructive testing model of meat, and output the corresponding predicted sensory score from the non-destructive testing model of meat.

[0017] In some embodiments, the sample meat includes: a first sample meat that has undergone freezing treatment, and a second sample meat that has not undergone freezing treatment. Attached Figure Description

[0018] 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. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0019] Figure 1a This is a flowchart illustrating a non-destructive testing method for beef according to an exemplary embodiment of the present invention. Figure 1b This is a flowchart illustrating the training method of a beef nondestructive testing model in an exemplary embodiment of the present invention. Figure 2 This is a schematic diagram of the modules of a non-destructive testing system for beef according to an exemplary embodiment of the present invention; Figure 3a This is a flowchart illustrating a nondestructive testing method in an exemplary embodiment of the present invention; Figure 3b This is a flowchart illustrating a training method for a nondestructive testing model in an exemplary embodiment of the present invention. Figure 4a This is a comparison chart of the prediction accuracy of the non-destructive testing model in a test embodiment of the present invention; Figure 4b This is a comparison chart of the prediction accuracy of the non-destructive testing model in another test embodiment of the present invention; Figure 5 This is a schematic diagram of a hyperspectral image of fish meat in an exemplary embodiment of the present invention; Figure 6 for Figure 5 The diagram shows a schematic representation of the spectral bands in the hyperspectral image. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.

[0022] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] In this document, "and / or" includes any and all combinations of one or more of the listed related items.

[0025] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0026] As used in this specification, the term "about" typically means + / -5% of the value, more typically + / -4% of the value, more typically + / -3% of the value, more typically + / -2% of the value, even more typically + / -1% of the value, and even more typically + / -0.5% of the value.

[0027] In this specification, certain embodiments may be disclosed in a range-bound format. It should be understood that this "range-bound" description is merely for convenience and brevity and should not be construed as a rigid limitation on the disclosed range. Therefore, the description of a range should be considered as having specifically disclosed all possible subranges and the individual numerical values ​​within those ranges. For example, a description of the range 1-6 should be considered as having specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and the individual numbers within those ranges, such as 1, 2, 3, 4, 5, and 6. This rule applies regardless of the breadth of the range.

[0028] In this paper, a hyperspectral image (HIS) is a type of image containing rich spectral information, capable of recording the reflectance or radiance intensity of each pixel across multiple consecutive spectral bands. Unlike ordinary color images (such as RGB images), which only contain spectral information in the red, green, and blue bands, hyperspectral images have a higher spectral resolution, covering hundreds or even thousands of narrow bands. This allows hyperspectral images to capture subtle spectral features of matter, achieving fine analysis that is impossible with traditional imaging. In this paper, a hyperspectral image can refer to an image with a spectral resolution of 10⁻⁶. -2 Spectral images in the order of λ.

[0029] In this article, "freezing" refers to lowering the temperature of food (such as beef, fish, tripe, etc.) to a set temperature (which can be understood as a comprehensive setting considering the type of meat, transportation time, and ambient temperature), thereby reducing the rate of biochemical reactions and microbial changes, and extending the shelf life of the food. For example, freezing refers to lowering the temperature of meat below its freezing point, causing at least part of its internal moisture to transform into ice crystals, thus achieving a solidified or frozen state. This process can effectively inhibit the growth of microorganisms, prevent spoilage, and facilitate long-term preservation and handling.

[0030] In this article, "food processing" can refer to one or more of the following cooking methods, such as frying, deep-frying, braising, boiling, stir-frying, roasting, etc. Alternatively, "food processing" can also refer to the industrial processing methods used in food processing plants to prepare cooked products from raw meat (or vegetables), such as pre-prepared vegetable processing, etc. Correspondingly, cooked products, such as meat products, refer to products that can be consumed by consumers or professional tasters.

[0031] This invention provides a method for non-destructive testing of the freshness of frozen meat, which can predict the sensory score of the final quality of meat (i.e., the quality of meat products after food processing).

[0032] It is worth noting that in existing technologies, hyperspectral imaging (an objective input data point) is often used to coarsely detect whether meat has undergone freezing (an objective output data point). In other words, traditional meat testing employs an objective-to-objective prediction approach. However, the applicant notes that with the rapid development of the frozen food chain, this testing method will struggle to meet the industry's demand for high-quality frozen meat.

[0033] In response, the applicant broke away from the traditional objective-to-objective prediction path in meat testing (i.e., hyperspectral image – was it frozen?), and proposed a groundbreaking objective-to-subjective prediction path, that is, directly predicting the final quality of meat from objective hyperspectral images. This prediction of final quality (evaluated through sensory scoring) can more accurately and comprehensively predict the overall impact of freezing treatment on meat quality.

[0034] In other words, this invention proposes a method that directly predicts the final quality of meat after cooking based solely on human subjective perception of meat products, using only hyperspectral image input. Furthermore, to achieve this objective-subjective prediction, this invention proposes a model training method based on a dual-relationship framework, which specifically includes: 1) Establish an objective-subjective prediction path by utilizing the connection between hyperspectral images, physicochemical information, and sensory scores (in other words, use hyperspectral images to record objective information about meat, and use the physicochemical information on them to connect the objective images with subjective sensory scores, so as to help establish an effective mapping relationship between hyperspectral images and sensory scores). 2) Use auxiliary data (taking beef as an example, auxiliary data can be drip loss rate) to help create a mapping relationship between hyperspectral images and sensory scores, so as to avoid the model having difficulty learning or establishing an effective mapping relationship between objective images and subjective scores.

[0035] Based on the above dual relationship, the present invention can establish an effective connection between hyperspectral images (objective) and sensory scores (subjective), and thus predict the difference between the taste of thawed meat after freezing and before freezing, simply by obtaining the hyperspectral image of the thawed meat.

[0036] Furthermore, this refined assessment of the impact of freezing on taste is of paramount importance to the development of current cold chain transportation.

[0037] For example, with breakthroughs in freezing technology, it has become possible to revive frozen fish, a phenomenon where frozen fish are restored to vital signs through specialized techniques. This phenomenon primarily relies on advanced freezing technologies, such as cloud freezing and microfreezing, which maximize the preservation of the fish's cellular structure, allowing it to regain its vitality after thawing.

[0038] Furthermore, microscopic observation reveals that fish, beef, pork, shrimp, and other ingredients treated with cloud freezing contain only scattered tiny ice crystals within their cells, exhibiting a cellular structure remarkably similar to that of fresh ingredients. This results in superior preservation of flavor, color, texture, and nutrients. Taking CP Group's Sapphire Blue Shrimp as an example, compared to traditional tunnel freezing and liquid nitrogen quick-freezing machines, shrimp frozen using cloud freezing technology, after cooking, are virtually indistinguishable from live shrimp in both appearance and taste, with all indicators far surpassing its competitors, setting a new benchmark for food preservation technology.

[0039] The objective-subjective prediction path proposed in this invention can achieve a refined scoring of the taste of thawed meat products, and at the same time help to achieve standardized testing for the current emerging cloud freezing and micro-freezing technologies, providing a reliable guarantee for cloud freezing and micro-freezing technologies.

[0040] The present invention will first use beef as an example to illustrate the non-destructive testing process provided by the present invention.

[0041] Example 1 See Figures 1a-1b As shown, the present invention provides a non-destructive testing method for beef, comprising the following steps: S101, Obtain the beef non-destructive testing model, wherein the beef non-destructive testing model is created using the following steps: S1011, acquire multiple hyperspectral images of multiple beef samples to generate a hyperspectral image set, wherein the beef samples include: a first beef sample that has undergone freezing treatment and a second beef sample that has not undergone freezing treatment; the hyperspectral image samples are provided with a first label and a second label, wherein the first label is used to mark volatile basic nitrogen information and the second label is used to mark at least one auxiliary physicochemical index; In this embodiment, the first label and the second label are model learning labels. By marking the key physicochemical indicators of the hyperspectral image, the AI ​​model can be guided to conduct in-depth learning of the core data in a focused manner, thereby reducing the interference of irrelevant data on model training to a certain extent. For example, in some embodiments, samples will be collected from at least 100 beef samples. Further, samples may be collected from at least 200, at least 500, or at least 1000 beef samples.

[0042] For example, in some embodiments, different freezing processes refer to different conditions such as freezing temperature, freezing time, and freezing cooling rate.

[0043] For example, in some embodiments, the sample beef that was not frozen typically refers to beef obtained directly after slaughter or beef that was stored at room temperature and did not spoil.

[0044] For example, in some embodiments, at least one of the auxiliary physicochemical indicators is one or more of the following: pH, color, and thiobarbituric acid reduction value.

[0045] S1012, Obtain first auxiliary data, the first auxiliary data including: drip loss rate of multiple samples of beef from the first sample beef; In some embodiments, drip loss rate is an indicator that measures the proportion of weight loss of a particular substance due to natural dripping water under specific conditions. The drip loss rate is calculated as follows: Drip loss rate (%) = (weight of drip loss / initial weight) × 100%; Weight loss due to dripping water = Initial weight - Weight after placement; In some embodiments, the drip loss rate can be obtained by one or more of the following test methods: natural hanging method, bagging method, dropper method, pressure filter paper method, and centrifugation method.

[0046] S1013, the multiple samples of beef are processed into multiple beef products, and at least two beef products are obtained through different food processing methods; S1014, record at least one sensory score given by the beef taster to each of the multiple beef products, and associate the sensory score with the hyperspectral image and / or the first auxiliary data; In some embodiments, tasters can conduct professional evaluations of beef’s appearance (such as color and texture), texture (tenderness and juiciness), and flavor (aroma and aftertaste) through observation, tasting, and other methods, and give corresponding sensory scores.

[0047] In some embodiments, the sensory score includes two or more of the following: color, odor, texture, viscoelasticity, broth after boiling (clear / turbid), and mouthfeel. Each of these has a sub-score, and the sum of the sub-scores is the final sensory score.

[0048] For example, in some embodiments, tasters need to have a keen sense of taste and smell, be able to accurately distinguish basic tastes such as sour, sweet, and salty as well as complex flavor layers (such as the differences in meat quality between different breeds of cattle), and have the ability to taste meat at national or regional tasting events or other industry-specific tasting events.

[0049] For example, with the optimal edible quality of beef after aging as 100 points, a scientific and reasonable sensory evaluation table can be developed to score the edible quality of thawed meat.

[0050] S1015, Generate a training set, the training set including multiple training sample groups, the training sample groups including: a hyperspectral image sample of a first sample of beef, and a first auxiliary data and sensory score associated with the hyperspectral image sample; a hyperspectral image sample of a second sample of beef, and a sensory score associated with the hyperspectral image sample. S1016, The training set is input into the multimodal fusion model. After the training process converges, the final multimodal fusion model is output as the beef non-destructive testing model. S102, Obtain at least one hyperspectral image of the beef to be tested; S103, input at least one of the hyperspectral images to be tested into the beef non-destructive testing model.

[0051] In some embodiments, the non-destructive testing model for beef outputs a frozen freshness value, which is the ratio of the sensory scores of the beef to be tested before and after freezing treatment.

[0052] For example, in some embodiments, the non-destructive testing model for beef can generate a corresponding predicted sensory score, and the sensory score of the beef to be tested before freezing can be obtained in advance through testing. Alternatively, in other embodiments, the sensory score of the beef to be tested before freezing can have a user-defined default value.

[0053] For example, in some embodiments, the training set—the hyperspectral image set—is obtained based on multiple beef samples, and an exemplary process for obtaining the hyperspectral images is as follows: Taking sample beef A as an example, sample beef A can refer to beef from the same batch (beef obtained from slaughter at the same time) or a piece of beef. First, sample beef A can be divided into at least two pieces of sample beef A1 and sample beef A2.

[0054] Further, at least one hyperspectral image P1 of the beef sample A1 is obtained, and then the drip loss rate D1 of the beef is obtained by testing a portion of the beef sample A1 (or, in some embodiments, the drip loss rate D1 of the beef sample that has not been frozen does not need to be tested), and the other portion of the beef sample is processed into food (such as steaming, boiling, stir-frying, etc.) to obtain beef product B1.

[0055] Meanwhile, sample beef A2 was frozen and then thawed, and at least one hyperspectral image P2 of the thawed beef was obtained. The drip loss rate D2 of the beef was obtained by testing a portion of sample beef A2, and the same food processing was performed on another portion of the beef to obtain beef product B2.

[0056] At least one beef taster will evaluate beef products B1 and B2 using sensory evaluation, and record at least one sensory score X1 and X2 for each product. A single sensory score for a beef product can be a characteristic value of multiple sensory scores given by various tasters; characteristic values ​​include the mean, median, etc.

[0057] Ultimately, the training set comprises multiple sets of sample data (or training sample sets) generated from multiple samples of beef. One set of sample data generated from sample beef A includes: Data prior to freezing: Hyperspectral image P1, drip loss rate D1, sensory score X1; Data after freezing: hyperspectral image P2, drip loss rate D2, sensory score X2.

[0058] Alternatively, in some embodiments, the data prior to freezing can be: hyperspectral image P1 and sensory score X1 (i.e., untested drip loss rate). Of course, in some embodiments, to ensure that the model can intelligently learn the mapping relationship between the hyperspectral image, drip loss rate, and sensory score, a default drip loss rate (such as the typical drip loss rate for that variety) can be set for samples where the drip loss rate was not tested.

[0059] Furthermore, in some embodiments, the same sample beef A can be divided into multiple samples beef A1, beef A2, ..., beef AN, and each sample beef A2... sample beef AN is subjected to different freezing treatment methods (e.g., different freezing temperatures, freezing times, freezing cooling rates, etc.). The same testing method is used to record the correlation data of the multiple samples beef (i.e., hyperspectral images before and after freezing, drip loss rates before and after freezing, and sensory scores before and after freezing).

[0060] Furthermore, in some embodiments, the same thawed sample of beef can be processed into multiple beef products using the same food processing method, and the sensory scores (such as the mean or median) of the multiple beef products can be recorded as the sensory scores corresponding to the sample of beef.

[0061] In some embodiments, meat products (such as beef products) can be comprehensively scored by referring to the standardized scoring system (such as appearance, taste, flavor, etc.) given in industry tasting events (such as beef tasting conferences) to select high-quality products.

[0062] In some embodiments, in S101, principal component analysis and multiple linear regression methods can be used to establish a hyperspectral-based fusion model to identify one or more auxiliary physicochemical information through hyperspectral analysis.

[0063] For example, in some embodiments, key physicochemical data can be identified from hyperspectral images through principal component analysis. For instance, principal component analysis can be performed on hyperspectral image samples to obtain the classification of the entire sample. Based on this, a class model corresponding to each type of sample is established. Then, the unknown sample is reclassified according to the class model, that is, the unknown sample is fitted with the class model of each sample to determine the category of the unknown sample. This can also be called soft independent pattern classification, thereby realizing the identification of physicochemical information in hyperspectral images.

[0064] In some embodiments, the non-destructive testing model for beef is further used to output a frozen-fresh value, which is the ratio of the sensory score of the sample meat (such as beef) before freezing to the sensory score after freezing. In this embodiment, the frozen-fresh value reflects the overall impact of freezing on the taste of meat; the smaller the frozen-fresh value, the smaller the overall difference in taste between frozen and fresh meat.

[0065] This embodiment proposes a method for evaluating meat freshness by correlating relatively objective hyperspectral images with relatively subjective sensory scores. By introducing the subjective evaluation of meat products by professional tasters, a more comprehensive and detailed assessment of meat freshness and changes in taste can be made, thus providing a suitable meat freshness detection solution for special freezing processes such as cloud freezing and micro-freezing.

[0066] In some embodiments, the specific process of adding a first label or a second label to a hyperspectral image is as follows: First, volatile basic nitrogen information (which can be used to assess bacterial contamination in meat) and auxiliary physicochemical indicators can be reflected in the hyperspectral image through different band information. Correspondingly, a second label can be set for the band reflecting color information, and a first label can be set for the band reflecting volatile basic nitrogen information. The labeled hyperspectral image sample and the first auxiliary data (preferably drip loss rate) are input into the multimodal fusion model, and the multimodal fusion model can actively learn the relationship between the combined parameters of the first label, the second label, and the first auxiliary data and the sensory score.

[0067] For example, in some embodiments, the wavelength of the band marked by the second tag is 550-650nm, and the wavelength of the band marked by the first tag is 400-500nm.

[0068] In this embodiment, color, volatile base information and drip loss rate in the hyperspectral image are preferably used as the main objective factors, and the relationship between these objective factors and subjective factors (sensory scores) is learned through model training.

[0069] Preferably, two or more hyperspectral images can be collected for the same beef sample. For example, in some embodiments, the hyperspectral image sample includes: a first hyperspectral image obtained from the beef sample, and at least one second hyperspectral image obtained by performing at least one image processing step on the first hyperspectral image.

[0070] The applicant noted that, for a large piece of beef, the properties of the meat may differ in different regions due to various reasons (such as different slaughtering methods, differences in the muscle properties of the beef itself, etc.). To improve the accuracy and reliability of the model, this invention also provides a sample optimization method based on multispectral image selection to guide the model to learn about beef with significant differences in property distribution as a whole, in order to output a comprehensive score.

[0071] In some embodiments, the image processing includes one or more of the following methods: cropping, masking.

[0072] In some embodiments, S1011 includes the step of: S10111, Identify the spectral reflectance data of at least two sub-spectral regions in the first hyperspectral image; S10112, Based on the spectral reflectance data, plot at least two target curves corresponding to the reflectance variation with wavelength; S10113, calculate the similarity index between at least two target curves; wherein, the similarity index is used to reflect the degree of similarity between the two target curves.

[0073] For example, in some embodiments, the similarity index can be the correlation coefficient between two curves, such as the Pearson correlation coefficient. Alternatively, in some embodiments, the similarity index can also be calculated using distance-based methods (such as weighted eigenvalue distance) or morphological analysis-based methods.

[0074] S10114, when the similarity index is less than the first set threshold (i.e., the band data of the two corresponding target regions are considered to be significantly different), at least one of the sub-spectral regions is selected as the target region. S10115, the target region is processed using the image processing method to obtain the second hyperspectral image, which does not contain the target region.

[0075] For example, in some embodiments, spectral reflectance data of sub-spectral region 1 and sub-spectral region 2 can be obtained for the first hyperspectral image. Target curve 1 and target curve 2 of sub-spectral region 1 and sub-spectral region 2 are calculated respectively. When the similarity index between target curve 1 and target curve 2 is low, it is considered that the difference between sub-spectral region 1 and sub-spectral region 2 is large. In this case, one of the sub-spectral regions can be randomly occluded to generate a new hyperspectral image (equivalent to a second hyperspectral image).

[0076] Preferably, in some embodiments, at least three sub-spectral regions are obtained, and a similarity index between any two sub-spectral regions is calculated. The sub-spectral region with the largest difference in similarity index from the other at least two sub-spectral regions is selected as the target region. For example, in some embodiments, when the similarity index between one sub-spectral region and the other at least two sub-spectral regions is less than a first preset threshold, it is selected as the target region.

[0077] In this embodiment, processing local data can improve the model's overall information about the hyperspectral image, that is, the overall quality of the beef, and avoid interference from local information.

[0078] In other words, this embodiment reduces the interference of local differences in beef quality on the overall result, which better meets consumers' overall focus on meat quality.

[0079] In some embodiments, beef can be any one of the following cuts: tenderloin, shank, brisket, ribeye, chuck, short ribs, and ribeye.

[0080] In some embodiments, the first auxiliary data further includes: chewability.

[0081] In some embodiments, the first auxiliary data further includes: cooking loss rate.

[0082] In some embodiments, the first auxiliary data further includes shear force.

[0083] It is understood that, in this embodiment, in order to achieve non-destructive testing of meat, a prediction path is provided to predict meat quality based on non-destructive testing data (i.e., hyperspectral images). Furthermore, to achieve non-destructive testing, this invention actually uses a combination of non-destructive and destructive testing data (such as dripping water loss) for training during the model training phase.

[0084] In this invention, non-destructive testing uses image data, while destructive testing uses non-image data (i.e., structured data). To achieve simultaneous learning and training of both image and structured data, this invention employs multimodal fusion technology. The key steps of the fusion process are illustrated below: 1. Data Preprocessing and Input Design Image processing: Image features can be extracted using convolutional layers of pre-trained models (such as VGG16 and ResNet). The input size can be uniformly set to a fixed resolution and then normalized.

[0085] Parameter processing: Numerical parameters such as non-image data are standardized (Z-score or Min-Max). Categorical variables can be encoded as one-hot vectors and compressed into vectors that match the dimensions of image features through fully connected layers.

[0086] In this way, image features and parameter vectors can be normalized (such as L2 normalization) to avoid training instability.

[0087] In other words, a dual-branch fusion structure can be adopted, specifically divided into an image branch and a parameter branch, to learn image data and non-image data simultaneously.

[0088] 2. Model architecture design: For a dual-branch fusion structure, fusion strategies can be employed for fusion. Specifically, one or more of the following fusion strategies can be used to fuse images and structured data: Concatenation: Concatenates the image feature vector and parameter vector along the channel dimension and inputs them into the classification / regression layer.

[0089] Weighted fusion: Dynamically assigns weights to images and parameters through an attention mechanism to enhance key features.

[0090] Cross-modal interaction: Use the Transformer module to implement feature interaction, such as using parameter vectors as contextual conditions for image features.

[0091] 3. Training and Optimization – One or more of the following methods can be used for optimization training: Joint training An end-to-end training approach is adopted, where the image branch and the parameter branch share the same loss function (such as cross-entropy or mean squared error), and the weights are updated synchronously through backpropagation.

[0092] Regularization methods When the amount of data in the parameter branches is small, Dropout and weight decay (L2 regularization) can be used to prevent overfitting.

[0093] Learning rate adjustment The image branch uses a lower learning rate, while the parameter branch can use a higher learning rate to balance the update speed of the two types of data.

[0094] The above methods can effectively integrate images and structured information, improving the model's performance in complex tasks.

[0095] Furthermore, to achieve prediction results with only an input image (i.e., avoiding the acquisition of lossy data), the parameter branch weights can be frozen after training, and a single image input can be implemented using any of the following methods: Zero-padding method: Input a vector of all zero parameters to keep the model structure unchanged; Mean substitution: Use the mean of the training set parameters as the default input value; Knowledge distillation optimization method: Construct a sub-model containing only image branches, and transfer multimodal knowledge to a single-modal model through a teacher-student framework.

[0096] It is understood that the model details given above are merely illustrative examples. Under the objective-subjective training concept proposed in this invention, the above model can be adaptively modified and adjusted according to different actual prediction objects.

[0097] See Figure 4a As shown, to verify the accuracy of the non-destructive testing method provided by the present invention, training model 1 (corresponding to the preferred non-destructive testing model of the present invention), training model 2, and training model 3 were used to test the sample accuracy. Among them, the overlap between the predicted sensory scores and the reference scores of training model 1 and 90% of the beef samples (or hyperspectral samples) was more than 90%. The reference score refers to the manual score of professional tasters.

[0098] See Figure 4bAs shown, to verify the accuracy of the non-destructive testing method provided by this invention on tripe, three other training models 1-3 were used to test the sample accuracy. It can be seen that, based on the proposed model training concept, by appropriately selecting physicochemical indicators, the sensory score predicted by the final model can highly overlap with the reference score, thus having guiding significance.

[0099] Example 2 See Figure 2 As shown, the present invention also provides a rapid and non-destructive testing system for the frozen freshness value of beef, comprising: The model module is used to obtain a non-destructive testing model for beef. This model is created using a model creation module, which includes: The image acquisition unit is used to acquire multiple hyperspectral image samples from multiple beef samples to generate a hyperspectral image set, wherein the beef samples include: a first beef sample that has undergone freezing treatment and a second beef sample that has not undergone freezing treatment; the hyperspectral image samples are provided with a first label and a second label, wherein the first label is used to mark volatile basic nitrogen information and the second label is used to mark at least one auxiliary physicochemical index; The data acquisition unit is used to acquire first auxiliary data, which includes: the drip loss rate of the first sample beef. The scoring and recording unit is used to record at least one sensory score given by at least one beef taster to the plurality of beef products; wherein, the plurality of beef samples are processed to obtain the plurality of beef products. A training set generation unit is used to generate a training set, which includes multiple training sample groups. The training sample groups include: a hyperspectral image sample of a first sample of beef, and a first auxiliary data and a sensory score associated with the hyperspectral image sample; and a hyperspectral image sample of a second sample of beef, and a sensory score associated with the hyperspectral image sample. The training unit is used to input the training set into the multimodal fusion model to train the corresponding beef non-destructive detection model. The image acquisition module is used to acquire at least one hyperspectral image of the beef to be tested; The output module is used to input at least one of the hyperspectral images to be tested into the beef non-destructive testing model, and the beef non-destructive testing model outputs a frozen fresh value, which is the ratio of the sensory scores of the beef to be tested before and after freezing treatment.

[0100] It is understood that the non-destructive testing system in this embodiment can implement any of the above methods or steps, which will not be elaborated here.

[0101] The preceding text used beef as an example to introduce a non-destructive testing method provided by this invention. Of course, the non-destructive testing method provided by this invention can also be applied to different types of meat. The following text uses general meat testing as an example to further explain the technical solution of this invention: Example 3 See Figures 3a-3b As shown, the present invention provides a non-destructive testing method for meat, comprising: S201, Obtain a non-destructive testing model for meat, wherein the non-destructive testing model for meat is trained using the following steps: S2011, obtain multiple hyperspectral images of multiple meat samples to generate a hyperspectral image set. Preferably, the hyperspectral image samples are provided with a first label and a second label. The first label is used to mark volatile basic nitrogen information, and the second label is used to mark at least one auxiliary physicochemical index. In some embodiments, the sample meat can all be undried meat, meaning that the model trained can also be applied to predict the cooking taste of fresh meat.

[0102] Preferably, in some embodiments, at least two samples of meat were subjected to different freezing treatments, and at least two samples of meat were not subjected to freezing treatment. Therefore, the resulting model can be used to predict the freshness of meat (such as frozen freshness value).

[0103] For example, in some embodiments, see Figure 5 , Figure 6 As shown, hyperspectral images of fresh live crucian carp, crucian carp frozen for 1 day, crucian carp frozen for 4 days, and crucian carp frozen for 16 days were obtained as part of the training sample set.

[0104] S2012, Obtain a type of auxiliary sample, wherein the type of auxiliary sample includes one or more of the following types of auxiliary data of the sample meat: drip loss rate, cooking loss rate, shear force, chewiness. In other words, in this embodiment, the collected non-destructive testing data (i.e., hyperspectral images) and destructive testing data (i.e., auxiliary data) are used to generate the subsequent training sample set.

[0105] S2013, The multiple meat samples are processed to obtain meat products; S2014, record at least one sensory score given by at least one taster to the meat product; For example, a single sample of meat can be used to prepare one or more meat products, and these meat products can be prepared using different food processing methods. The final recorded (or associated) sensory score of a single sample of meat can be the average score obtained from multiple meat products.

[0106] S2015, Generate a training set, which includes multiple training sample groups, including hyperspectral image samples before and after freezing, and the hyperspectral image samples are respectively associated with a type of auxiliary sample and sensory score; S2016, The training set is input into the multimodal fusion model for model training; after the training process converges, the corresponding non-destructive meat detection model is obtained. S202, Obtain at least one hyperspectral image of the meat sample to be tested. S203, at least one of the hyperspectral images to be tested is input into the non-destructive testing model for meat, and the non-destructive testing model for meat outputs the corresponding predicted sensory score.

[0107] In some embodiments, the type of meat can be any of the following: beef, fish, shrimp, tripe, lamb, etc.

[0108] In some embodiments, S2016 includes the step: S20161, The hyperspectral image sample is obtained; S20162, at least one group of auxiliary physicochemical indicators is screened from the hyperspectral sample; S20163, at least one auxiliary data group is selected from the aforementioned auxiliary samples; S20164, Generate multiple condition combinations based on the volatile basic nitrogen information, at least one combination of auxiliary physicochemical indicators, and at least one combination of auxiliary data; S20165, The multimodal fusion model is used to learn the training set by combining multiple conditions to obtain multiple non-destructive detection models.

[0109] In some embodiments, auxiliary physicochemical indicators may also be one or more of the following: pH, color, thiobarbituric acid reduction value, protein content, etc.

[0110] Among these, pH refers to the acidity or alkalinity of the meat, color refers to the color and luster of the meat, and volatile basic nitrogen (TVB-N) is the alkaline nitrogenous substance produced by the decomposition of proteins during spoilage due to the action of enzymes and bacteria, resulting in ammonia and amines. Thiobarbituric acid reduction value refers to the content of malondialdehyde (MDA) in a sample determined by the thiobarbituric acid reactive substances method (TBARS), used to evaluate the antioxidant capacity of the sample.

[0111] Different storage methods after animal slaughter have a significant impact on its quality. This invention preferably selects the above-mentioned physicochemical information from hyperspectral images to serve as an indicator for judging the freshness of meat under different storage conditions (such as refrigeration and freezing) and different storage times.

[0112] Furthermore, auxiliary data may include one or more of the following: cooking loss rate, drip loss rate, shear force, and chewiness.

[0113] Cooking loss rate refers to the reduction in quality that occurs during the processing of raw meat into cooked meat due to moisture loss during cooking. Drip loss rate refers to the proportion of water lost from the muscle of meat after slaughter. Shear force refers to the relative displacement deformation of the cross-section of meat along the direction of the lateral external force (i.e., the force perpendicular to the plane of action). Chewability refers to the mechanical properties of food during the chewing process.

[0114] The above-mentioned auxiliary data are mainly used to evaluate the edible quality of meat after slaughter. The cooking loss rate and drip loss rate are used to judge the water-holding capacity of the meat, while the shear force and chewiness are used to judge the tenderness of the meat.

[0115] In some embodiments, the multiple hyperspectral image samples include: a first hyperspectral image sample obtained for a first sample meat, and a second hyperspectral image sample obtained for a second sample meat; wherein, the second sample meat is the first sample meat or meat from the same batch of the first sample meat after freezing treatment; correspondingly, the sensory evaluation includes: a first sensory evaluation of the first meat product obtained by the taster from the first sample meat, and a second sensory evaluation of the second meat product obtained by the taster from the second sample meat.

[0116] In some embodiments, correspondingly, S20165 includes the step: A first training subset and a second training subset are selected from the training set, and the first training subset is trained with the multimodal fusion model under at least two combination conditions, thereby obtaining multiple first models. Calculate the accuracy of at least two first predictions for the at least two first models; A second model that meets a first screening rule is selected from multiple first models. The first screening rule requires that the first prediction accuracy of the second model is greater than a first accuracy threshold. The second model is optimized using the second training subset; Calculate the accuracy of at least two second predictions for the at least two second models; When the second prediction accuracy is greater than the second accuracy threshold, at least one of the corresponding second models will be output as non-destructive testing models.

[0117] In this embodiment, for different types of meat, such as beef, mutton, fish, tripe, etc., different parameter combinations can be selected from physicochemical information and auxiliary data to obtain a targeted non-destructive testing model.

[0118] Understandably, as the sample size increases or the scope of model application expands, the model can be further subdivided to meet application needs.

[0119] To address this, multiple detection models can be trained simultaneously, and the accuracy preferences of these models when facing different types of meat to be tested can be recorded at the same time. Based on historical preferences, suitable models can then be recommended to users.

[0120] Preferably, in some embodiments, there may be two training phases in the initial application phase. The application process is explained below using mutton as an example: First training phase: m1 sets of parameter combinations are selected from physicochemical information and auxiliary data. Models are trained for each of the m1 parameter combinations to obtain m1 training results. If the model training process converges, the training result is considered the initial model; if the model training process fails to converge, the training result is invalid. Therefore, m2 (m2≤m1) initial models (equivalent to the first model) can be obtained from the m1 parameter combinations. The first prediction accuracy of each of the m2 initial models is calculated; if the first prediction accuracy is greater than the first prediction accuracy, the initial model is selected as the second model. The first training subset is used in this training phase.

[0121] Second training phase: The second training subset is used to update and train the second model, and finally m2 optimized second models are obtained as non-destructive detection models.

[0122] Preferably, in some embodiments, the accuracy of the second model will be further verified, and m3 (m3≤m2) models with higher accuracy will be selected from the second model as the final non-destructive testing model.

[0123] Model testing / application phase: Obtain hyperspectral images of multiple different mutton samples, input them into multiple non-destructive testing models, record the prediction results of each non-destructive testing model, and then evaluate the accuracy of each non-destructive testing model on different hyperspectral images.

[0124] In some embodiments, S202 includes: (1) Input multiple hyperspectral images to be tested into at least two non-destructive testing models and obtain at least two sets of non-destructive testing results, wherein the hyperspectral images to be tested are marked with category information, and the category information includes: meat variety and / or food processing method; For example, taking mutton as an example, its breeds can be divided into goat meat, sheep meat, and wild sheep meat. Furthermore, based on the geographical location of its breeding, it can be categorized as follows: Albas goat meat, Sunite goat meat, Ujumqin goat meat, Yanchi Tan goat meat, Shaanxi Hengshan goat meat, Xinjiang Hotan goat, Ordos fine-wool goat meat, Guangdong Leizhou goat, Henan Huai goat, Sichuan Nanjiang yellow goat, Qinghai Tibetan goat, Guizhou Qianbei Ma goat, Dazu black goat, Ningxia Yanchi Tan goat, Xinjiang Bashibai goat, Xinjiang Bayinbuluke black-headed goat, etc.

[0125] (2) Calculate the third prediction accuracy of at least two sets of nondestructive testing results; For example, in some embodiments, the prediction results can be recorded and evaluated during the application process.

[0126] (3) When the difference between two of the third prediction accuracies is greater than the third accuracy threshold (i.e., the same non-destructive testing model has a large difference in the recognition accuracy of different hyperspectral images), the category information of the hyperspectral image with the higher third prediction accuracy is identified as the update information; wherein, the non-destructive testing model records at least one weight level, and the weight level is associated with category information; (4) Use the updated information to update the recommended weights of the nondestructive testing model; The steps for updating the recommendation weights include: (41) Obtain the set value of the weight level corresponding to the updated information; (42) Increase the set value according to the set rules.

[0127] For example, in some embodiments, when a nondestructive testing model shows significant differences in accuracy in identifying different types of meat products (i.e., different varieties or different food processing methods), the application information of the nondestructive testing model will be updated. The application information describes the nondestructive testing model's preference for different varieties or food processing methods.

[0128] For example, multiple hyperspectral images to be tested may include: a first set of spectra corresponding to the breed category of goat meat, and a second set of spectra corresponding to the breed category of sheep meat. When it is identified that the prediction accuracy of one of the nondestructive testing models for the first set of spectra is significantly higher than that for the second set of spectra, the category information of the first set of spectra, namely goat meat, can be obtained. Correspondingly, the nondestructive testing record has or is generated based on the category information a weight level. In some embodiments, the initial weight level may have a default setting value.

[0129] Furthermore, if the new hyperspectral image continues to be input into the nondestructive testing model, and the nondestructive model continues to show a significant advantage in goat meat, then the corresponding weight level can be updated, such as by incrementing the level by 1. Understandably, the rules for setting the level update can be defined by the user.

[0130] In other words, the present invention also provides a model update and optimization method that can update the model in accordance with the detailed requirements of actual application scenarios, so as to recommend more suitable models or prediction results to users.

[0131] For example, in some embodiments, when there are multiple non-destructive testing models, the category information of the hyperspectral image to be tested can be obtained, and the detection results output by the multiple non-destructive testing models can be displayed sequentially according to the weight level of the non-destructive testing model for the category information.

[0132] Furthermore, in some embodiments, when the weight level for a certain category (such as a certain variety or a certain food processing method) is greater than the set weight level, it is preferable to obtain a new training set for that category and use the new training set to update the non-destructive testing model. In some embodiments, at least one hyperspectral image to be measured includes: At least one first hyperspectral image covering a first area region of the meat to be tested, and at least one second hyperspectral image covering a second area region of the meat to be tested; wherein the first area region is larger than the second area region.

[0133] In some embodiments, S202 includes: Acquire at least one first hyperspectral image covering a first area region of the meat to be tested; Identify the spectral reflectance data of at least three sub-regions in the first hyperspectral image to be measured; Based on the spectral reflectance data, at least three target curves corresponding to the reflectance variation with wavelength are plotted respectively; Calculate the similarity index between any two of the at least three target curves; When the similarity index between one of the target curves and at least two other target curves is less than the first similarity threshold, the sub-region to be tested corresponding to the target curve is identified as the region of interest. A new second hyperspectral image to be measured is obtained for the region of interest; For example, in some embodiments, a new hyperspectral image can be taken for the region of focus.

[0134] Alternatively, in other embodiments, the region of interest in the first hyperspectral image to be measured can be masked to obtain a new second hyperspectral image to be measured.

[0135] For example, in some embodiments, it can be determined whether the region of interest has learning value in order to select whether the information of the region of interest needs to be included in the second hyperspectral image to be measured.

[0136] The first and second hyperspectral images to be tested are input together into the non-destructive testing model.

[0137] It should be noted that the proportions of muscle, fat, and connective tissue may vary in different regions of meat, resulting in different spectral reflectance characteristics. For example, differences in the chemical composition of fat and muscle will be reflected in the spectral absorption characteristics of specific wavelengths (such as shortwave infrared). Color differences (such as variations in myoglobin content) may also be reflected in the spectral response. For instance, different parts of beef will exhibit variations in color depth due to differences in myoglobin concentration, thus affecting the reflectance of hyperspectral images.

[0138] Furthermore, edible quality indicators (such as tenderness, juiciness, and marbling) are not uniformly distributed in meat products. Hyperspectral imaging can identify localized variations in these indicators through spectral differences, such as differences in muscle fiber arrangement or fat deposition areas. Simultaneously, if adulterants (such as foreign objects or non-meat components) are present in localized areas, their spectral characteristics will show a significant contrast with normal areas.

[0139] In other words, hyperspectral images can contain rich physicochemical information, and the information characteristics at different locations on the image may differ significantly. To address this, the present invention provides a sample selection method that guides a model to comprehensively identify the overall quality of meat. Therefore, the resulting non-destructive testing model can be used to identify the freshness of meat in most scenarios.

[0140] In some embodiments, the nondestructive testing model is further used to output a frozen freshness value, which is the ratio of the sensory score of the sample meat before freezing to the sensory score after freezing.

[0141] Preferably, in some embodiments, the predicted scores for the model (such as frozen-fresh value, sensory score) can also be presented in a visual form. For example, a hyperspectral image labeled with the predicted score can be output. As another example, a large hyperspectral image can also be divided into regions for prediction, and then each region of the hyperspectral image can be labeled with a predicted score.

[0142] Understandably, users can set their own specific scoring rules. For example, if the standard score is set to 100, a score of 90 or above can usually be considered as indicating that the meat is fresh and tastes good.

[0143] Example 4 The present invention also provides a non-destructive testing system for meat, comprising: The model module is used to obtain a non-destructive testing model for meat. This meat non-destructive testing model is created using a model creation module, which includes: The first sample acquisition unit is used to acquire multiple hyperspectral image samples of multiple meat samples to generate a hyperspectral image set. The hyperspectral image samples are provided with a first label and a second label. The first label is used to mark volatile basic nitrogen information, and the second label is used to mark at least one auxiliary physicochemical index. The second sample acquisition unit is used to acquire a type of auxiliary sample, which includes at least one type of auxiliary data of the sample meat, namely: drip loss rate, cooking loss rate, shear force, and chewiness. The scoring and recording unit is used to process the multiple meat samples to obtain meat products and record at least one sensory score given by the tasters to the meat products. A training set generation unit is used to generate a training set, which includes multiple training sample groups, each of which includes hyperspectral image samples and is associated with a type of auxiliary sample information and a sensory score. The training unit is used to input the training set into the multimodal fusion model for model training. After the training process converges, the corresponding non-destructive meat detection model is obtained. The output module is used to acquire at least one hyperspectral image of the meat to be tested, input the at least one hyperspectral image to the non-destructive testing model of meat, and output the corresponding predicted sensory score from the non-destructive testing model of meat.

[0144] In some embodiments, the nondestructive testing model is further used to output a frozen freshness value, which is the ratio of the sensory score of the sample meat before freezing to the sensory score after freezing.

[0145] It is understood that the non-destructive testing of meat described in this invention is only an example, and the concept of model training or model application provided by this invention can also be applied to the field of non-destructive testing of other foods, such as vegetables.

[0146] It is understood that the system in this embodiment can also implement the steps of any one or more of the above methods, which will not be repeated here.

[0147] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0148] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0149] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A non-destructive testing method for meat, characterized in that, include: S201, Obtain a non-destructive testing model for meat, wherein the non-destructive testing model is trained using the following steps: S2011, acquire multiple hyperspectral images of multiple meat samples to generate a hyperspectral image set, wherein the hyperspectral image samples are provided with a first label and a second label, the first label is used to mark volatile basic nitrogen information, and the second label is used to mark at least one auxiliary physicochemical index; S2012, Obtain an auxiliary sample, wherein the auxiliary sample includes at least one of the following auxiliary data of the sample meat: drip loss rate, cooking loss rate, shear force, and chewiness. S2013, The multiple meat samples are processed to obtain meat products; S2014, record at least one sensory rating given by the taster to the meat product; S2015, Generate a training set, the training set including multiple training sample groups, the training sample groups including: hyperspectral map samples, and the hyperspectral map samples are respectively associated with a type of auxiliary sample information and sensory scores; S2016, The training set is input into the multimodal fusion model for model training. After the training process converges, the corresponding non-destructive meat detection model is obtained. S202, acquire at least one hyperspectral image of the meat to be tested, input the at least one hyperspectral image of the meat to be tested into the non-destructive testing model of the meat, and output the corresponding predicted sensory score from the non-destructive testing model of the meat.

2. The method according to claim 1, characterized in that, S2016 includes the following steps: S20161, The hyperspectral image sample is obtained; S20162, at least one group of auxiliary physicochemical indicators is screened from the hyperspectral sample; S20163, at least one auxiliary data group is selected from the aforementioned auxiliary samples; S20164, Generate multiple condition combinations based on the volatile basic nitrogen information, at least one combination of auxiliary physicochemical indicators, and at least one combination of auxiliary data; S20165, The multimodal fusion model is used to learn the training set by combining multiple conditions to obtain multiple non-destructive detection models for meat.

3. The method according to claim 2, characterized in that, S20165 includes the following steps: A first training subset and a second training subset are selected from the training set, and the first training subset is trained with the multimodal fusion model under at least two combination conditions, thereby obtaining multiple first models. Calculate the accuracy of at least two first predictions for the at least two first models; A second model that meets a first screening rule is selected from multiple first models, wherein the first screening rule requires that the first prediction accuracy of the second model is greater than a first accuracy threshold. The second model is optimized using the second training subset; Calculate the accuracy of at least two second predictions for the at least two second models; When the second prediction accuracy is higher than the second accuracy threshold, at least one of the corresponding second models will be output as the non-destructive testing model for meat.

4. The method according to claim 2, characterized in that, S202 includes: Multiple hyperspectral images to be tested are input into at least two non-destructive testing models for meat, and at least two sets of non-destructive testing results are obtained accordingly. The hyperspectral images to be tested are marked with category information, which includes: meat variety and / or food processing method. Calculate the third prediction accuracy for at least two sets of nondestructive testing results; When the difference between two of the third prediction accuracies is greater than the third accuracy threshold, the category information of the hyperspectral image with the higher third prediction accuracy is identified as the update information; wherein, the non-destructive meat detection model records at least one weight level, and the weight level is associated with category information; The updated information is used to update the recommended weights of the nondestructive testing model; The steps for updating the recommendation weights include: Obtain the set value of the weight level corresponding to the updated information; Increase the set value according to the set rules.

5. The method according to claim 2, characterized in that, At least one hyperspectral image to be measured includes: At least one first hyperspectral image covering a first area region of the meat to be tested, and at least one second hyperspectral image covering a second area region of the meat to be tested; wherein the first area region is larger than the second area region.

6. The method according to claim 2, characterized in that, S202 includes: Acquire at least one first hyperspectral image covering a first area region of the meat to be tested; Identify the spectral reflectance data of at least three sub-regions in the first hyperspectral image to be measured; Based on the spectral reflectance data, at least three target curves corresponding to the reflectance variation with wavelength are plotted respectively; Calculate the similarity index between any two of the at least three target curves; When the similarity index between one of the target curves and at least two other target curves is less than the first similarity threshold, the sub-region to be tested corresponding to the target curve is identified as the region of interest. A new second hyperspectral image to be measured is obtained for the region of interest; The first and second hyperspectral images to be tested are input together into the non-destructive testing model for meat.

7. The method according to claim 1, characterized in that, The sample meat includes: a first sample meat that has undergone freezing treatment, and a second sample meat that has not undergone freezing treatment.

8. The method according to claim 7, characterized in that, The non-destructive testing model is also used to output a frozen freshness value, which is the ratio of the sensory score of the sample meat before freezing to the sensory score after freezing.

9. A non-destructive testing system for meat, characterized in that, include: The model module is used to obtain a non-destructive testing model for meat. This meat non-destructive testing model is created using a model creation module, which includes: The first sample acquisition unit is used to acquire multiple hyperspectral image samples of multiple meat samples to generate a hyperspectral image set. The hyperspectral image samples are provided with a first label and a second label. The first label is used to mark volatile basic nitrogen information, and the second label is used to mark at least one auxiliary physicochemical index. The second sample acquisition unit is used to acquire a type of auxiliary sample, which includes at least one type of auxiliary data of the sample meat, namely: drip loss rate, cooking loss rate, shear force, and chewiness. The scoring and recording unit is used to process the multiple meat samples to obtain meat products and record at least one sensory score given by the tasters to the meat products. A training set generation unit is used to generate a training set, which includes multiple training sample groups, each of which includes hyperspectral image samples and is associated with a type of auxiliary sample information and a sensory score. The training unit is used to input the training set into the multimodal fusion model for model training. After the training process converges, the corresponding non-destructive meat detection model is obtained. The output module is used to acquire at least one hyperspectral image of the meat to be tested, input the at least one hyperspectral image to the non-destructive testing model of meat, and output the corresponding predicted sensory score from the non-destructive testing model of meat.

10. The system according to claim 9, characterized in that, The sample meat includes: a first sample meat that has undergone freezing treatment, and a second sample meat that has not undergone freezing treatment.

Images