Method for determining marbling of meat

Abstract

During the method for determining the meat and fat components that determine the marbling of a meat slice, and for classifying a meat slice based on its visual properties, a detailed image recording of at least "full HD" resolution is taken of the meat slice to be examined, so that the meat slice can be completely distinguished from a background of a different color than the colors identifiable on the meat slice, the image recording is recorded on a data medium, by processing the image recording, by transforming the colors to grayscale, normalizing them, and filtering them, then determining basic parameters normalized by noise, and additional gradient parameters using a gradient operator, mak¬ ing a prediction from the parameters using support vector machine method and correct it with a quadratic polynomial, rounding the resulting value to an integer and outputting it as a marbling class value.

Description

[0001] M ethod for determining marbling of meat

[0002] The invention relates to a method for determining the marbling of meat and for classifying a meat slice based on its visual properties.

[0003] M arbling is a parameter that quantifies the structure of meat and fat tissues, thereby predicting the expected enjoyment value and organoleptic properties of the meat during consumption. For this purpose, several methods have been developed, the common feature of which is that the meat and fat parts are separated in the meat slice to be examined, and a numerical value is determined based on their quantity and ratio, with which the meat slice is classified in accordance with the classification known and used in the field. A traditional solution is the visual examination of the meat slice with the naked eye, in which the ratio, location, appearance of the meat and fat are observed, and what is seen is evaluated according to professional criteria. This is addressed, for example, in the manual AUS-M EAT Limited (2021 ): Handbook of Australian beef processing - The AUS-M EAT language (page 30), where there are reference cards similar to the photographs shown in the attached drawing, to the pattern of which the judging persons com pare the surface of the meat slice or the cutting surface.

[0004] Patent specification WO 2014050168 Al describes a method in which an image is taken of a meat slice, a rib eye section (the meaty, fatty, juicy part of the beef rib with a nice big piece of fat in the middle) is determined and selected in the meat slice, the selected section is separated into muscle and marbled section, a ratio of the area of the digital image of the fat area and the area of the digital image of the muscle is calculated, and then a marbling fineness index is calculated from the image of the marbled section by dividing the sum of the boundary lengths of the marbled sections by the square root of the area of the rib eye section, and a marbling value is determined based on the fat area ratio and the fineness index. A shortcoming of the proposed solution can be considered that the applied determination method only allows a rudimentary and general classification, which in many cases is not suitable for comparison with the results of methods used in the given field.

[0005] Patent specification US 6891961 B2 describes an image analysis system for meat classification and meat quality prediction, in which an image of a cross-section of a rib eye is captured by a camera with a wedge-shaped camera housing that facilitates insertion into the rib incision. The image capture portion of the system includes a camera with a flash for uniform illumination. Since the camera is positioned to capture the cross-section of the rib eye at an angle to accommodate the wedge shape of the camera housing, the camera housing is also provided with various alignment devices that allow the user to make adjustments to obtain a uniform image. After the image has been digitally captured or captured and converted to a digital image, image analysis is performed on the digital image to determine parameters such as percentage leanness, total rib eye area, total fat area, total lean area, percentage marbling, fat thickness near the rib eye, and other parameters that can also be used to indicate and classify the marbling of the meat slice. The shortcoming of this proposed solution is that the imaging technique and environment used may introduce initial uncertainties and distortions into the process that, during subsequent calculation and determination, presumably unintentionally, lead to a result that differs from the real one.

[0006] A study "Spektralis kepfeldolgozas merestechnikai eredmenyei, alkalmazasa husok marvanyozottsaga- nak, erlelesi idejenek jellemzesere" (Measurement results of spectral image processing, application to the characterization of meat marbling and maturation time), (NIR Club, 18.04.2013) of Gergely Eder and Dr. Ferenc Firtha shows a method that can be learned, among others, for determining meat marbling, the shortcoming of which is that it is a high hardware and time-consuming, expensive method that can be performed primarily in laboratory conditions, by determining significant wavelengths, thus it cannot be used in an industrial environment, or it can only be used with difficulty.

[0007] Meat marbling determination solutions known from the state of the art do not provide a simple, inexpensive, reliable method that can be used in an industrial environment, which provides the same or nearly the same result even repeatedly. Our invention aims to satisfy this need.

[0008] We have recognized that regardless of the calculation and determination methods used, it is essential to ensure the properties of the data source serving as the basis for the calculations and the classification determination. In this case, this means that the starting source, i.e. image, required for grading meat marbling must be obtained in a manner that provides essentially the same result.

[0009] The set task has been solved by a method having the features according to claim 1 . Some advantageous embodiments of the method have been formulated in dependent claims.

[0010] The invention will be described in more detail below with the help of an exemplary embodiment of the method, with reference to the attached drawing, in which

[0011] Figure 1 shows an original color (RGB) image of a meat slice subject to the method,

[0012] Figure 2 shows a normalized, transformed image of the meat slice subject to the method,

[0013] Figure 3 shows an image of the meat slice subject to the method after noise removal,

[0014] Figure 4 shows an inverted gradient image of the meat slice subject to the method after noise removal, including edges,

[0015] Figure 5 shows an estimation of marbling using an SVM model, and

[0016] Figure 6 shows an estimation of marbling using a two-stage model. In the method, which is presented only as a preferred embodiment, the marbling of meat is determined for classifying a meat slice based on its visual properties.

[0017] In this specification the expression "meat" means a cut of meat prepared for culinary processing. In order to ensure the reliability of the method, the method is preferably performed on freshly cut cut or cuts of meat, by determining the surface pattern of the cut of meat based on a color image taken of the cut of meat, and the marbling class is determined from these data. The reference data for the marbling class is provided by expert sensory evaluation, as is currently the case for meat classification. The method is detailed, for example, in the aforementioned AUS-M EAT Ltd manual. The color images produced provide the same view as viewed by the experts, therefore the presented method actually models expert decision-making, with a complex mathematical formulation of the properties.

[0018] The first main stage of the presented method is the preparation of the image to be taken of the cut of meat. Correct appearance of colors is essential for the analysis of the images, therefore lighting and environment should be standardized.

[0019] After cutting the meat slice to be examined - which can be one or more meat slices - it can be determined visually that it contains essentially red meat tissue parts and essentially white fat tissue parts. In order to clearly recognize and distinguish these, the meat slice must be placed in front of a colored background, which color can be clearly distinguished from the two colors of the meat slice mentioned, which are important to us, both by eye and, preferably, by digital image recording. To this end, in the example presented, the meat slice is placed on a test surface providing a blue background, which is a plastic tray in the exemplary method. The color of the tray is blue when illuminated with a light source with a correlated color temperature of D65 (according to the international standard D: daylight, 65: 6500 K) and measured with an instrument of the type ColorLite sph850. The illumination closely approximates the spectral distribution of natural sunlight. When choosing the tray, make sure that its size is sufficiently larger than the size of the meat slice to be examined so that its blue color provides a background surrounding the meat slice in the image to be created.

[0020] Reference colors are used for the test, Table 1 shows the blue background color of the images to be taken in different color spaces, using the CIE (ISO 1 1664-4:2008(E) / CIE S 014-4 / E2007) color system, also showing values converted to other color systems: Table 1

[0021] In order to set the background color of the image to be taken and to ensure the comparability of further determinations, after recording, it is ensured that the lighting does not shift the color balance. The lighting can be natural white or cold white light, not so-called "warm white", because yellow shifts the colors towards red. During the method, preferably LED lighting with a color temperature of 3500 - 4000 K is used. In order to avoid reflection and glare, indirect and homogeneous lighting should be used. In order to avoid glare and similar optical disturbances, the background and the surface of the meat must not be wet, i.e. they must not contain water droplets or moisture. This is checked before taking the image, and if necessary, moisture removal in any known way should be done.

[0022] The second stage of the presented method is the image creation itself. For the correct evaluation of the observed marbling pattern, a nearly constant level of detail, i.e. resolution, is required. The recording itself can be done with any color camera, which provides at least a so-called “full HD”, i.e. pixel image size of 1920 * 1080. The image size and the area recorded by the camera together provide a constant resolution and detail. The size of the tray used in the presented implementation of the method is 38 * 27 cm, the resolution of the full HD image recorded by the camera or resized to full HD is 0.025 cm / pixel, i.e. 0.25 mm / pixel. When creating the image, we ensure that the optical axis of the camera is essentially perpendicular to the plane of the meat slice to be examined, and the image includes the entire tray area. An important aspect is that, unlike known methods, we do not use any other illumination besides the predefined and set illumination, e.g. we do not use a flash either. The image taken is recorded on a data storage medium, such as an SD card, in accordance with the technical characteristics of the camera used, which is handled in the computing device used in a third stage presented below in a manner known and customary in the field, but of course known wired or wireless communication can also be used.

[0023] The third stage of the presented method is the processing of the image taken. In the presented method a software "Scilab" (version 2024.1 .0) known to a person skilled in the art is used with the addition of an IPCV (Image Processing and Computer Vision) toolbox, version 4.5.0.

[0024] The image recorded by the camera and recorded on the data carrier is first resized so that its smallest side (width or height) corresponds to the full HD size of 1920 * 1080 pixel. The image can be larger, but not smaller. This results in an effective resolution of at least 0.25 mm / pixel. The resolution value changes if the image ratio deviates from the Full HD ratio. This resizing is performed using a known bilinear approximation. Such a method can be found, for example, in the paper Press, William H.; Teu- kolsky, Saul A.; Vetterling, William T.; Flannery, Brian P. (1992). "Numerical recipes in C: the art of scientific computing" (New York, NY, USA: Cambridge University Press, pp. 123-128. ISBN 0-521 -431 08- 5.).

[0025] The next step is the transform ation of the colors. Using the red, green, and blue color channels R, G, and B of the color image shown in Figure 1 , a gray image L is created in a manner known in the field of image processing:

[0026] L = R + G - 2* B

[0027] The purpose of the color transformation is to separate the red meat, white fat and blue background. The data of the R, G, B color channels are 8-bit integers in the range of 0-255, the colors and their transformed values are listed in Table 2:

[0028] Table 2

[0029] Following the above transformation, in the next step, the obtained values are normalized to the range 0-255 using the following formula: L — min L

[0030] LN- 255 max L — min L

[0031] In the normalized image shown in Figure 2, the blue background is given an intensity value of 0 and appears as a black background. The normalized values of the colors shown in Table 2 are: 255 for the red meat surface, 1 70 for the white fat parts, and 0 for the blue background. The background can be clearly separated, while the points of the meat surface fall in the range of approximately 170 - 255. As a result of the normalization, the gray image automatically adjusts to the red color of the meat and the blue color of the background.

[0032] In this example, the well-known Otsu threshold method is used to separate the meat from the background and to select the computational area, see for example Dmitriy Csetverikov: "Basic algorithms for digital image analysis", Eotvos Lorand University, Faculty of Informatics. This creates a binary mask with which we select which part of the image is analyzed. To remove possible noise, an “open” operator is applied using a square mask of 9* 9 pixel. The mask covers an area of 2.25 mm * 2.25 mm on the meat surface. Successive erosion and dilation remove small-scale noise and create a uniform analyzable surface.

[0033] This image shown in Figure 3 is used to calculate the following four parameters:

[0034] - normalized image variance

[0035] - normalized image contrast value

[0036] - normalized image energy value

[0037] - normalized image entropy value.

[0038] In the next step, a search for edges in the image is carried out using a “gradient” operator. An edge is the difference between the selected point and its surroundings. Figure 4 shows the corrected image and the gradient. The latter is shown inverted to show the edges. Using this image, we calculate four more parameters:

[0039] - gradient image variance

[0040] - gradient image contrast value

[0041] - gradient image energy value

[0042] - gradient image entropy value.

[0043] In this way, eight parameters are calculated in the normalized transformed image (LN) and the gradient image containing the edges. These are variance, contrast K, energy E and entropy S. The latter three can be calculated using the following formulas: where i is the intensity value in the range 0-255, and P(i) is the frequency of its occurrence. The purpose of all parameters is to specify the density of the fat tissue pattern on the surface of the meat slice.

[0044] In the fourth stage of the method, marbling is determined. To estimate the marbling, the eight data mentioned above is used, all of which are calculated only for the meat surface, not for the background:

[0045] From the data, a prediction is made using the Support Vector Machine (SVM) method, see for example “Chang, Chih-Chung and Lin, Chih-Jen (201 1 ) LIBSVM : A library for support vector machines, ACM Transactions on Intelligent Systems and Technology, 2(3): 27:1-27:27”, which is corrected with a quadratic polynomial (PM ). In the presented example, the calculation is performed using a PM + SVM method.

[0046] The first step is to apply the SVM model. The model was created with a "radial" kernel function, with a gamma (y) weighting factor of 0.8. In the next step, the density value calculated for the SVM meat slice surface is adjusted to the 1 -dimensional expert scale. The graph in Figure 5 shows the reference assessment given by the experts, the “Expert assessment” and the value calculated by the model, the "SVM calculated density". The correlation for this data series is close, r = 0.9759 (Pearson's linear correlation).

[0047] After the parabolic transformation of the values given by the SVM model, the statistical indicators shown in Table 3 are obtained during the qualitative evaluation of the two-step regression.

[0048] Table 3:

[0049] The coefficient of determination indicates a very close relationship, no significant autocorrelation can be expected, and the estimation error, RM SE, has a value of 0.160. The autocorrelation can be further improved with a large number of additional measurements, but based on the collected data, the model is already usable. The value of marbling is usually indicated by an integer. The value of the estimation error 0.160 indicates that the uncertainty of the model is below the rounding limit.

[0050] The correction equation used here is:

[0051] M = a + SVM(b + c SVM) where M is the estimated marbling, SVM is the density estimated from the measured data using the SVM method, and a, b, and c are constants determined during calibration.

[0052] The estimated and actual values using the two-stage model (PM +SVM ) are shown in Figure 6. In the Figure, the rounded marbling class values marked on the “ PM + SVM result” curve all fall on the diagonal, i.e. they are the same as the reference value indicated on the horizontal axis, “ Expert assessment”. This calculated value is rounded to an integer and is output as a marbling class value, i.e. shown, displayed or recorded as data according to its intended use.

Claims

Claims1. M ethod for determining marbling of meat, the method comprising- in an image recording stage, obtaining an image of a meat slice to be examined,- storing the obtained image in a data storage medium,- in an image processing stage, processing the image stored in the data storage medium by a processing unit comprising a processor and a memory in communication with the processor in such a way as to determine the meat and fat components determining the marbling of the meat slice,- in a determination stage, classifying the marbling value of the meat slice into a marbling class based on the determined ratio of the meat and fat components of the meat slice, wherein the method comprising the further steps: prior to the image recording stage, in a preparation stage: arranging the meat slice to be examined in front of a background of a different color from the red meat tissue parts and the white fat tissue parts of the meat slice, such that in such a way that the background color surrounds the meat slice to be examined, illuminating the background and the meat to be examined indirectly and homogeneously with natural white or cold white light without the use of additional lighting, checking a presence of water droplets and moisture on the background and the surface of the meat, and removing the water droplets and moisture present there, in the image recording stage following the preparation stage: adjusting a relative position of a camera used to obtain the image and the meat slice to be examined so that the optical axis of the camera is perpendicular or nearly perpendicular to the plane of the meat slice to be examined, and an image to be taken includes the entire area of the background, obtaining an at least so-called "full HD" color image i.e. a color image of a size of 1920 * 1080 pixels and storing the image data on a data storage medium, in the image processing stage following the image capture stage, in relation to the image data stored on the data storage medium, checking whether the size of the image created, stored and processed is 1920 * 1080 pixels, if not, then resizing the image to 1920 * 1080 pixels, and if so, continuing the processing of the image,during image processing, transforming the colors of the color image to separate the red meat, white fat and blue background parts using the following formula:L = R + G — 2 - B where L is the transformed value, is the red, G is the green, and B is the blue color component value; converting the color image containing the transformed colors to a grayscale image, normalizing the resulting grayscale image values to the range 0-255 using the following formula:where L is the transformed intensity of the meat slice surface, LNis the normalized intensity, selecting a part of the image to be analyzed to separate the meat slice from the background, performing noise removal on the part of the image to be analyzed, based on the denoised image, determining following four basic parameters:- normalized image variance- normalized image contrast value- normalized image energy value- normalized image entropy value, searching for edges in the image using a “gradient” operator, based on the image containing the edges, determining additional four parameters:- gradient image variance- gradient image contrast value- gradient image energy value- gradient image entropy value using the eight defined parameters, determining the density of fatty tissue pattern on the surface of the meat slice,In the determination stage following the processing stage: by calculating the eight defined parameters on the surface of the meat slice, determining the marbling value of the meat slice by making a prediction from the parameters using a Support Vector M achine (SVM) method corrected with a quadratic polynomial,rounding the calculated value to an integer and outputting the value as a marbling class value.

2. The method according to claim 1 , characterized in that a freshly cut meat slice is selected in the preparatory stage.

3. The method according to claim 1 or 2, characterized by selecting a blue tray is as the background in the preparatory stage, the size of which exceeds the size of the meat slice in all directions.

4. The method according to any one of claims 1 -3, characterized by checking the absence of optical interference in the preparatory stage.

5. The method according to any one of claims 1 -4, characterized by using the known Otsu threshold method in the image processing stage to separate the meat from the background, with which a binary mask is created.

6. The method according to any one of claims 1 -5, characterized by using an “open” operator of a square mask of 9* 9 pixel in the image processing stage to remove noise.

Images