KR102285112B1 - Non-destructive measurement system of meat yield of beef carcasses using digital image analysis - Google Patents

Abstract

Provided is a beef carcass meat yield prediction model construction system comprising: an image collection step of collecting carcass images of various states by using a color camera; an image processing step of extracting a feature by processing the images collected by the prior step and including a background removal step, a key point detection step, and a feature extraction step; and a model construction step of constructing a prediction model with a multiple linear regression (MLR) method by using the selected feature among the extracted features. According to the present invention, real-time carcass evaluation is possible by applying the system to a slaughter line.

Description

Non-destructive measurement system of meat yield of beef carcasses using digital image analysis

The present invention relates to a non-destructive cattle carcass meat measurement system using digital image analysis, and more particularly, to a model building system capable of predicting the meat quantity of cattle carcass through color images by combining multivariate statistics and image analysis. will be.

Beef is the third most consumed meat in the world after pigs and chickens, accounting for about 25% of global consumption. This consumption is likely to continue to increase over the next few decades, given the growing population and growing wealth of developing countries. Therefore, the recent beef distribution network and related laws and regulatory agencies are very interested in measuring and evaluating the quality of distributed beef in advance.

Recently, beef production and consumption are rapidly increasing, and for this reason, problems of measuring and monitoring the quality of beef produced in advance are emerging. Grading is a method of measuring the quality of beef before distribution, and through quality evaluation, certification for safe livestock products and the grade and price of carcasses are determined. As a result, grading has a direct economic impact on farmers and meat distributors. Therefore, it is necessary to increase the accuracy of grading in the slaughterhouse, and there is a need to conduct accurate grading before the slaughtered carcases are distributed.

For consumers, the quality of beef is very important, and as a result, it acts as a determinant of product selection. Currently, the beef quality grade is conducted by quality assessors who have been trained in the slaughterhouse, and the amount of meat is determined after being boned by hand. However, this process consumes a lot of time and money, and subjective factors may be involved as the evaluation standards and skill levels are different for each judge and slaughterhouse.

According to the American Meat Scientist Association (AMSA), it is known that maturity, marbling, texture and firmness of meat, meat color and fat color are factors that determine the quality of carcasses. Maturity is determined by the age of the animal and is related to the degree of hardening of the vertebral bones and cartilage. Maturity affects meat color and texture, which in turn affects overall meat quality. Marbling is so named because intramuscular fat occurs as small dots like marble patterns. Marbling is associated with juiciness and quality of meat. The maturity and marbling of carcasses are the main evaluation criteria for the quality evaluation of cattle carcasses, and the final quality is evaluated by evaluating meat color, texture, and fat characteristics together. According to a previous study (Ricardo et al., 2016), muscle quality and fat content were found to be important quality indicators in the grading of carcasses. The quality of muscle refers to the amount of muscle determined by the structure of the muscle or the ratio of muscle to bone, and fat refers to the degree of deposition inside or outside of fat.

The evaluation of bovine conductors is usually carried out through a systematic method of determining quality based on standard diagrams. The economic value of a carcass is generally determined by the saleable meat yield (SMY%), lean meat percentage (LMP), and weight of large cuts (loin, ribs, hind legs, tenderloin, etc.). LMP is one of the most commonly used carcass yield quality indicators and serves as an important criterion for classifying carcasses. This LMP is a key parameter for carcass classification in most European countries and around the world, and is a numerical value expressed as a percentage of meat weight to carcass weight. Several studies have documented LMP as lean meat yield (LMY), defined as the ratio of carcase weight to meat plus fat and some bones, depending on the market (Craigie et al., 2012).

LMP is being measured in real time at the slaughterhouse using a variety of equipment and technologies. Depending on the degree of automation in the slaughterhouse, these equipments include manual and semi-automatic types such as Fat-O-Meat'er and UltraFom (Carometec A/S, Herlev, Denmark), and automatic models include AutoFom (Carometec A/S, Herlev). , Denmark), VCS2000, VBS2000 (from E + V Technology GmbH, Oranienburg, Germany), and BCC (Frontmatec Group, Roskilde, Belgium). In addition, according to the operation method of the device, it is divided into light-reflecting measurement technology, ultrasound, electromagnetic wave, or vision technology-based devices. Most of these equipment and technologies are currently used on meat processing lines, and other equipment is utilized for a variety of purposes. Examples include techniques such as fat content determination, age determination, country of origin determination, and carcase quality determination.

The above-mentioned equipment and technology are very helpful in predicting the LMP of insulators, but there are still improvements that need to be improved. In some cases, these devices have disadvantages in that they take time to measure, are complicated to use or are expensive to introduce, and have low accuracy as shown in Table 1 below. The low accuracy of these systems may be influenced by problems with image acquisition methods, feature extraction methods, lack of statistical prediction methods, or sample data sets used for model development. However, the prediction of conductor weight and shape similarity between conductors of these pre-developed systems showed mostly accurate results because they are related to the morphological characteristics of conductors.

Location (System) sample number Quality parameter Performance ^* Reference Australia (BBC-3) 427 Pistol cut (kg) R ² = 0.967
RMSE = 0.91 (Jakob et al., 2019) Denmark (BCC-2) 3459 Saleable meat (%) R ² = 0.70 SEP = 1.31 (Borggaard, Madsen, & Thodberg, 1996b) USA (VIAScan) 240 Hot carcass weight (kg) R ² = 0.75 RSD = 1.08 (Cannell et al., 1999) Germany (VBS2000) 301 Fat class R ² = 0.75 SEP = 1.20 (Sonnichsen et al., 1998)

* RMSE: root-mean-square error; SEP: standard error of prediction; RSD: relative standard deviation

The present invention relates to lean meat percentage (LMP), hot carcass weight (HCW) and large divisions of 10 parts (loin, sirloin, loin tip, pork loin, forelimb, rump, scallop, brisket, scallop, To provide a simple and non-destructive image analysis-based system for predicting ribs) weight.

The present invention provides an image collecting step of obtaining images of conductors in various states using a color camera; processing the image collected in the above step to extract features, the image processing step including a background removal step, a key point detection step and a feature extraction step; and a model building step of constructing a predictive model using a multiple linear regression (MLR) method using the selected features among the extracted features.

In addition, the image collecting step is characterized in that it is a step of obtaining a conductor inward direction image and a conductor dorsal direction image of a right conductor or a left conductor in the state of a temperature body or a cold conductor.

In addition, in the background removal step, the final mask image is obtained through a morphological image processing process for the image generated by removing the blue screen region, improving the conductor contrast through color statistics, and combining the threshold setting and edge detection. This is characterized in that the background is removed.

In addition, the key point detection step selects a point that can be the starting point of classification for each part of the conductor as a key point, selects a model conductor image, designates a master image with key point labeling, and when a new image is collected, the master image It is characterized in that it is a step of mapping the key point position to a new image by matching with .

In addition, the feature extraction step is a step of extracting variables used in the predictive model, characterized in that it is a step of extracting features using the mask image obtained in the background removal step and the key point obtained in the key point detection step.

In addition, the model building step is characterized by using at least one of a feature selection cleaning (FSC) algorithm and an SFS sequential feature selection) algorithm as an algorithm for selecting a feature to be used for modeling among the extracted features.

The present invention combines multivariate statistics and image analysis to predict the amount of meat in cattle carcases through color images. The performance of the model built through the characteristics selected through image analysis and MLR ^{was confirmed that R 2} was greater than 0.86 for all parameters. In particular, the prediction performance for LMP ^{showed high prediction accuracy with R 2} = 0.89 and RMSE = 1.99%. This performance corresponds to relatively good performance compared to most VIA systems such as VBS2000 and VIAScan. Therefore, real-time carcass evaluation is possible by applying the system according to the present invention to a slaughter line.

1 shows an image collection image according to an angle of a conductor.
2 shows the correlation between LMP089 and LMP based on reference data.
3 shows a process for detecting and removing a blue screen using the LAB color space.
4 shows the contribution of each band to the final image, a plot of regression coefficients used for conductor detection, and color regression results.
5 shows a process of obtaining a mask image through threshold setting, edge detection combination, and morphological image processing.
6 shows a master image in which key points are projected onto a new conductor image through KAZE feature matching and affine transformation.
7 shows a conductor image matching process for key point estimation.
8 is a visualization of the features (RP, A, B, C, and E features) extracted from Image No. 1 (the inner part of the conductor).
9 is a visualization of features (RP, A, B, and C features) extracted from image No. 2 (dorsal part of the conductor).
10 shows a schematic diagram from image acquisition to final yield prediction model construction.
11 shows a histogram plot of the weight, age and LMP of the carcass collected for validation.
12 shows the importance of selected parameters in the LMP prediction model.
13 shows the visualization and construction of selected properties for predicting LMP using the likelihood image of the thermobody state.
14 is a regression plot (top) and residual plots (bottom) generated from LMP regression modeling using three foreign parameter variables (right) for all features (left), SFS-selected features (middle), and SFS-selected features. is shown.
15 shows a regression plot showing the predictive performance of each model for carcase gross weight and other major cuts.

All technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art, unless otherwise stated. Also throughout this specification and claims, unless stated otherwise, the term comprise, comprises, comprising is meant to include the recited object, step or group of objects, and steps, and any other It is not used in the sense of excluding an object, step, or group of objects or groups of steps.

Prior to describing the present invention in detail below, it is to be understood that the terminology used herein is for the purpose of describing specific embodiments and is not intended to limit the scope of the present invention, which is limited only by the appended claims. shall.

On the other hand, various embodiments of the present invention may be combined with any other embodiments unless clearly indicated to the contrary. Any feature indicated as particularly preferred or advantageous may be combined with any other feature and features indicated as preferred or advantageous. Hereinafter, embodiments of the present invention and effects thereof will be described with reference to the accompanying drawings.

According to an embodiment of the present invention, a system for constructing a meat quantity prediction model of cattle carcass is a system for non-destructively predicting the meat quantity of cattle carcass through a color image by combining multivariate statistics and image analysis, comprising: an image collection step; image processing step; and a model building step through multivariate statistical analysis;

The image collection step is a step of obtaining images of conductors in various states by using a color camera, and includes a right conductor or a left conductor in a hot or cold conductor state, an image in the direction of the inside of the conductor (an image at an angle of 0˚) and an image of the dorsal side of the conductor ( 90˚ angle image).

In acquiring images, camera calibration may be continued. Camera calibration is a process of estimating internal or external parameters of a camera. The internal parameters refer to camera internal characteristics such as focal length, distortion, and image center, and external parameters refer to the position and orientation of the camera. Camera calibration continues until the re-projection error for the shooting area is less than 0.3 pixels.

The image processing step is a step of extracting features by processing the collected images, and includes a background removal step, a key point detection step, and a feature extraction step.

The background removal step is a step for extracting the silhouette of the conductor image. After removing the blue screen region and improving the conductor contrast through color statistics, morphological image processing for the image generated by combining threshold setting and edge detection This is the step of removing the background by obtaining the final mask image through the process.

In the key point detection step, first, a point that can be the starting point of classification for each part of the conductor is selected as a key point, a model conductor image is selected, a master image with key point labeling is designated, and then a new image (analysis target) is selected. image) is collected, it is a step of mapping the key point position to a new image by matching it with the master image.

The feature extraction step is a step of extracting variables used in the predictive model, and is a step of extracting features using the mask image obtained in the background removal step and the key point obtained in the key point detection step. The extracted features include features extracted based on the characteristic areas of the mask, center, and contour, features extracted based on 16 cross-sections, centers and contours, and features extracted based on key points. In addition, as external variables, the temperature body weight (HCW, nit:Kg), the age (months) of the slaughtered cattle, sex, etc. can be added.

The step of constructing a model through the multivariate statistical analysis is a step of constructing a predictive model using a multiple linear regression (MLR) method using selected features among the extracted features. It has better performance than a predictive model built using Gaussian Process Regression (GPR) using all features.

An algorithm for selecting a feature to be used for modeling among the extracted features uses at least one of a feature selection cleaning (FSC) algorithm and a sequential feature selection (SFS) algorithm.

<Example>

(1) Sample data collection

Data collection was carried out for 3.5 years (January 2017 - June 2020) at the slaughterhouse of the National Institute of Livestock Science in Jeonju. For this example, measurements were made for 140 Korean cattle raised in different regions by gender, age, and weight.

(2) Image data collection

For image measurement of carcases, two-part carcasses in the temperature body state after slaughter and the cold carcass state in which the carcases were cooled at 2°C for 24 hours were acquired. The quality evaluation of the conductor was conducted by a quality assessor after taking an image of the conductor in the cold conductor state. Image acquisition was performed for both right and left conductors. A color camera (Matrix Vision GmbH, Oppenweiler, Germany) was used as the camera used for filming, and it was controlled by being connected to a computer via USB3. The camera sensor used mvBlueFOX3-2059C combined with a wide-angle lens to acquire an image of the front side of the conductor.

A total of two images were taken for each divided conductor (left conductor, right conductor), and as shown in Fig. 1, an image at an angle of 0˚ (inward direction of the conductor_medial section image) and an image at an angle of 90˚ (the dorsal side of the conductor) Direction_Dorsal section). It was rotated using a servo motor (Ezi-Servo plusR, Fastech, Gyeonggi-do, South Korea) for 90˚ angle image acquisition, and was controlled by the same computer. The camera and motor were synchronized in one program and used for image shooting.

A custom UI was developed in MATLAB software (MathWorks Inc., Natick, MA, USA) using the "Genlcam" adapter from EMVA (European Machine Vision Association) to capture images from the camera in a specific instance. In addition, control was carried out by combining the developed UI with a software development kit provided by the motor manufacturer to capture a small conductor image at a specific angle. In order to obtain an optimal image during shooting, the exposure time for shooting was set to 25 ms. Afterwards, the acquired file was saved as a file in portable network graphic (PNG) format, and then stored in a predefined directory for performing the image processing step.

Camera calibration is the process of estimating internal or external parameters of a camera. Intrinsic parameter variables refer to camera internal characteristics such as focal length, distortion, and image center. The extrinsic parameters refer to the position and orientation of the camera. Lens barrel distortion (due to the use of wide-angle lenses) that can affect system performance after camera calibration has been removed. A standard checkboard method was used for camera calibration in this study. Calibration was continued until the re-projection error for the shooting area was less than 0.3 pixels.

(3) Reference data collection

Reference data for model development (carcass weight, gender, age, large division weight, etc.) were collected on site before and after workers dismantled the carcass. On the other hand, LMP was calculated by applying Equations (1) and (2), which are EU standard formulas for LMP, to the obtained reference data.

Equation (1):

Equation (2):

As can be seen in Equation (1), LMP089 is calculated based on the weight of the five main divisions, whereas LMP is calculated considering all the meat dismantled from the whole carcass. In the conductor data set used in this example, LMP089 and LMP showed a high correlation as shown in FIG. 2 . In this example, LMP was used instead of LMP089 as the final reference data.

(4) Image processing

Feature extraction was performed after image processing was performed on the collected carcass images. In this chapter, we divided the analysis into three parts focusing on background removal, key point detection, and shape extraction.

(4.1) Background detection (background removal)

The background removal process aims to extract the silhouette of the carcass image. However, the background was not uniform due to the spatial limitation of the slaughterhouse where the study was conducted as shown in FIG. 1 . For this reason, the background could not be removed by simple methods such as setting color intensity and color thresholding. To solve the background removal problem, a step-by-step approach to this problem was attempted using various image processing techniques. The first was to remove the blue screen area, the second to improve the conductor contrast through color statistics, and the third to create a mask image by combining threshold setting, edge detection and morphological operation. .

(4.1.1) Blue screen detection

The blue screen can be easily detected by using the b* (blue - yellow value) value in the LAB color space. In the histogram of the b* value of the acquired image, a local minimum was used as a cut-off value to create an image from which the blue screen was removed ( FIG. 3 left ). After the blue screen of the image was detected, the values of all pixels corresponding to the blue screen were set to 0. The masking image and the result of masking the conductor are shown in FIG. 3 .

(4.1.2) Conductor Contrast Enhancement Using Color Statistics

For the following process, 500 carcass photos in the data set were selected and image segmentation was performed manually. At this time, the background was labeled as 0, and the pixel corresponding to the conductor was labeled as 1. Image pixels were converted into four color spaces (LAB, HSV, YCBCR, XYZ). A simple linear regression analysis was then performed between the labeled pixels and a matrix of 15 color variables to obtain the regression coefficients used as the conductor and background detection criteria in the conductor image. In addition, the contribution of each color band to the final prediction was analyzed to select the color channel with the highest correlation with the conductor. As a result, in the analysis, the red channel of RGB showed the highest correlation (0.84% importance score), followed by the y value of the xyz channel, the A* value of the LAB color space, the x value of xyz, and the S value of the HSV space. high was detected. The remaining items showed less than 10% of the importance score. The silhouette of the conductor image was predicted using the selected five color bands. A result of contrast enhancement based on color is shown in FIG. 4 .

(4.1.3) Combining Threshold Setting and Edge Detection

After detecting the foreground through color statistics, the generated image was transformed into a binary image, and then the image was transformed using OTSU thresholding. Also, the binary image produced by color statistics was edge-detected by Canny's edge algorithm. Then, the binary image generated by the edge detection method and the threshold image generated by the OTSU algorithm were synthesized. Thereafter, a morphological image processing process (image closing, image area opening, image filling, image filtering) was performed on the generated image, and a final mask image was produced through this process (FIG. 5).

(4.2) Key point detection

In this example, a selected point in the carcass was defined as a key point based on previous studies, Korean carcass dissection guidelines, and experiences obtained in the modeling process. A total of 16 key points were selected (FIG. 6). The key point algorithm was conducted under the basic assumption that the general shape of the conductor is similar to that of other conductors, and that parts (eg, ribs, legs, spine, etc.) can be easily distinguished from other conductors as well. In this embodiment, the image descriptor algorithm such as the speeded up robust features (SURF) algorithm, the scale-invariant feature transform (SIFT), and the KAZE feature algorithm are used to find the common part between the two images in the images of two different conductors. made it possible The KAZE algorithm used in this embodiment uses nonlinear diffusion filtering, which effectively reduces image noise and maintains the boundary of objects, so that functional positions can be clearly distinguished compared to other algorithms. In this embodiment, a random sample consensus (RANSAC) algorithm is used to additionally filter the actual robust match between the two images.

In this embodiment, a model conductor image is selected (FIG. 7), 16 key-point labels are attached, and this is designated as a master image. When a new image with no key point specified enters the analysis process, it is mapped with the master image generated by the KAZE and RANSAC algorithms and converted into a T matrix in the process. The T matrix maps key-point positions from the master image to the new image using Equation (3) below.

Equation (3):

where [u, v] is the coordinates of the new key point in the new image, and [x, y] is the original coordinate of the key point in the "master" image. The performance of the key point algorithm was evaluated using 30 images compared with the corresponding predicted positions of the algorithm after manual labeling (Table 2, FIG. 6). Performance criteria used for evaluation were coefficient of determination, R ² , and root meat square error (RMSE).

Although the results of the key-point detection algorithm used in this example showed high results (R ² =0.9, RMSE = 104 pixel, 157 mm), there were still significant errors. The cause of this error is judged to be caused by a combination of a human error made by a human during labeling and an area that does not match during image conversion. Therefore, it is judged that the improvement of these errors can improve the quality of the extracted features, and consequently the performance of the carcass butchery prediction model.

Mean Median standard dev R ² 0.8961 0.9072 0.0733 RMSE (pixels) 123.96 133.33 45.31 RMSE (mm) 187.14 201.28 67.93

(4.3) feature extraction

The feature extracted through the feature extraction process of this embodiment means a variable used in the predictive model, and was extracted using a conductor mask image and predefined key points.

Finally, the extracted features were classified into six arbitrarily defined main categories: RP, A, B, C, D, and E. The detailed description of the extracted features is shown in Table 3.

Features Number Remarks Based on region properties of mask, centroid and outline RP features Area One Area of carcass mask bounding box 2 Bounding box dimensions Perimeter One perimeter of mask Major and minor axes 2 Major and minor axes of carcass mask Extent One Ratio of mask area to bounding box Orientation One Banking angle of carcass A feature 4 Distance from centroid to point of intersection between outline and lines connecting centroid to bounding box vertices Based on 16 cross-sections, centroid and outline B features 16 From 16 segments of carcass along its length (i.e., carcass breadth at different sections) C features 32 Distance from centroid to point of intersection between outline and each of the 16 section line D features 1024 Inter-point distance between points of intersection between outline and 32 sections Based on key points E features 256 Inter-point distances between 16 detected key points TOTAL 1340

The RP features included information on the area of the masked conductor, the size of the bounding box, perimeter, major and minor axes, extent and direction, and were extracted from the properties of the masked area. RP consists of a total of 8 features.

Feature A was defined as the distance from the mask center to the intersection between the conductor contour and the line from the center to the vertex of the bounding box. Finally, feature A was classified into four features.

For B feature extraction, the bounding box surrounding the conductor was subdivided into a total of 16 identical segments. The length of the line at the edge of every segment within the boundary of the conductor contour was considered as the B feature. Therefore, a total of 16 features B were extracted.

Feature C is the distance between the center point and feature B, and a total of 32 C features were extracted.

To extract feature D, a total of 32 points were acquired through the conductor outline and feature C. The distance between the obtained combinations of points was defined as feature D. For feature D, a total of 32 ² (1024) numbers were extracted. However, in the case of feature D, overlapping features occurred repeatedly, and then were removed through feature cleaning in the model improvement process.

Feature E was derived from 16 key points obtained using the feature extraction algorithm described above. The distance between the extracted combinations of 16 points was defined as the feature E. As a result, a total of 16 ² (256) features were extracted for feature E. Feature E also showed a tendency for some elements to overlap each other.

Finally, all features were extracted from image 1 (inner part of the conductor) and 1340 features were extracted, and RP, A, B, and C features were extracted from image 2 (dorsal part of the conductor) to generate a total of 60 features. Features extracted from each image can be confirmed in FIGS. 8 and 9 . A total of 1,400 features extracted from the two images were used for initial model development (Table 3, Table 4).

image number Number of features Image 1, central camera (0°), from carcass medial section 1340 Image 9, central camera (90°), from carcass dorsal section 60 Total features 1400

Additionally, we tested the performance improvement of the model by adding three external variables that were not extracted from the image. Some of the existing commercialized VIA systems also incorporate external variables into the model, and by including external variables in this study, it helps to facilitate comparison with the existing VIA system. External variables added during model development were set as body weight (HCW, unit: Kg), age (months) of slaughtered cattle, and sex.

(5) Model development

In general, multivariate statistical analysis is useful for finding patterns in large-scale data that humans cannot find. This technology has recently been widely applied in pattern recognition, data search and classification problems. In this embodiment, multivariate data analysis was utilized to find patterns in features extracted through image processing, and a model for estimating the conductor yield parameters was developed. The final goal of the modeling process is to develop a simple regression model that predicts the carcass slaughter rate with good performance using minimal features.

(5.1) Development of an initial model using all features

In the case of the initial model, the model was built using all 1400 features extracted from images in two directions. However, in the case of the initial model, Gaussian process regression (GPR) was used instead of the simple regression algorithm in the modeling process because of the many features compared with the number of samples. The GPR model is a nonparametric kernel-based probabilistic model that approximates a set of random variables with a consistent Gaussian distribution. A GPR regression analysis model was evaluated to quantify its performance in predicting carcass yield parameters.

(5.2) Feature Selection

Since the number of features extracted from the image data is large in the initial model, there is a possibility that the predictor matrix contains noise or unnecessary or redundant features. In this embodiment, feature selection cleaning (FSC) and sequential feature selection (SFS) algorithms are applied to all parameters in order to reduce the noise effect and find features related to key points. In the case of the FSC algorithm, it serves to remove constants and correlated features. On the other hand, the SFS algorithm plays a role in finding the linear combination with the highest correlation with the response. In this study, both FCS and SFS algorithms were implemented. As a result, several characteristics were detected for final model development.

(5.3) final model

After feature selection, a final regression model for the selected features was developed. In the final model, both the GPR method and the MLR method could be tested during model development. MLR is also known as multiple ordinary least squares regression, and this analysis method generally fails to predict high-dimensional data and highly correlated data. However, the final model was able to apply MLR analysis when developing the final model because the number of features used by feature selection was reduced compared to the initial model.

In the final model development process, 70% of the total data was used for calibration and cross-validation, and the remaining 30% of the sample was used to test the final model.

In addition, some conductors in the data set were classified as outliers because they were erroneous or not properly detected, and were excluded from the model building process to generate reliable models. In this case, the outlier was identified by a robust multivariate dispersion algorithm.

<Experimental example>

(1) model evaluation

In the process of testing the model developed in this study, it was confirmed which conductor among the left and right conductors and the conductors in the temperature and cold conductor states produced the best predictive model. Also, when developing the initial model, the performance of the model developed using 1340 features extracted from image 1 and the model developed using 1400 features extracted from two images (images 1 and 2) were evaluated. Moreover, three external factors (age, gender, weight) were included in the development of the new model, and the obtained results were compared with the previous model.

To test the performance of individual models, goodness-of-fit criteria including ^{RMSE and coefficient of determination (R 2 ) were adopted for the calibration set and validation set.} The RMSE and R ² values were derived by equations (4) and (5).

Equation (4):

Here, 'r' is the residual obtained by subtracting the predicted value from the reference data, and 'n' is the number of samples.

Equation (5):

와

각각 reference 값과 예측된 값을 의미한다.

Wow

Refers to the reference value and the predicted value, respectively.

A summary of the entire process from data collection to final yield prediction model construction is shown in FIG. 10 .

<Experiment result>

(1) Sample data

A total of 140 small carcasses were used for performance verification, and a data set was constructed with 38 castrates, 40 males, and 62 females. The weight range of the thermobody was measured in the range of 239-598 kg, the age was 15-207 months, and the calculated LMP range was 48-82%. Details of the conductors are summarized in Table 5 and Figure 11. According to the histogram plot of FIG. 11 , a normal distribution for the temperature body and LMP except for age can be confirmed. According to FAO (1991), the typical slaughter age of cattle is known to be less than 80 months.

Sex Number HCW (Kg) Age (months) LMP (%) Min - Max Mean SD Min - Max Mean SD Min - Max Mean SD Castrated 38 299.5 - 597.9 455.40 74.11 21 - 35 30.29 2.97 49.20 - 71.43 58.62 4.401 Male 40 257.0 - 560.2 424.01 76.17 15 - 77 28.07 10.04 60.17 - 81.64 69.51 4.262 Female 62 238.8 - 523.1 376.97 33.55 20 - 207 64.51 33.55 47.89 - 70.45 60.05 4.567 Total 140 238.8 - 597.9 411.67 77.44 15 - 207 44.80 28.99 47.89 - 81.64 62.37 6.361

(2) model accuracy and model comparison

In most analyzes, the LMP yield parameters were used to quantify the best performing conductor state (temperature body, cold conductor) and orientation (internal side image, dorsal image), including the number of images required for model development, and external parameters. The final comparison was made by comparing the model built using all the extracted features (1400 pieces) with the model built through the selected features.

(2.1) Temperature body vs cold conductor

As a result of the prediction model, the features extracted from the image of the cooled conductor showed slightly higher LMP prediction accuracy than the temperature body (Tables 7 and 8). However, since this difference is not statistically significant, both models can be used for estimating the fleshing ratio. However, when applying the system to an actual slaughterhouse, it is appropriate to analyze the temperature body to obtain immediate carcass quality. Therefore, in this experimental example, it was decided to focus on the temperature body data for analysis.

(2.2) Left Conductor vs Right Conductor

Table 6 is a table comparing the results of the LMP prediction model constructed using the left and right conductors of the temperature body. In this comparison table, there was no significant difference between the two models of the left and right conductors. However, in this experimental example, it was decided to use the right conductor image. In general, in Korea, bisection is performed by attaching a tail to the left conductor, because there is a possibility that the model may become complicated in the feature extraction process during image processing. Subsequent analysis was conducted on the right conductor in the thermobody state.

left side right side R ² cal. RMSEC (%) R ² pred. RMSEP (%) R ² cal. RMSEC (%) R ² pred. RMSEP (%) Select. feat. 0.92 1.84 0.88 2.03 0.92 1.67 0.89 1.99 Select. feat + Ext. 0.90 2.00 0.88 2.13 0.91 1.83 0.91 1.99 all feat 0.94 1.39 0.84 2.12 0.89 2.45 0.84 2.95 Average 0.92 1.74 0.87 2.09 0.91 1.98 0.88 2.31

(2.3) single image vs two images

Table 7 is a table comparing the performance of the LMP prediction model developed using a single image and two images. The model developed by analyzing the two images showed better performance than the model based on one image, and ^{it was confirmed that the R 2} value had an average of 10% or more and the RMSE was about 40% different. This is thought to be due to the fact that the two image-based models have the characteristics of giving accurate information about the LMP measurement of the conductor. Since the model using two images showed better predictive performance, a model was built based on the two images for additional analysis.

2 images 1 image R ² cal. RMSEC (%) R ² pred. RMSEP (%) R ² cal. RMSEC (%) R ² pred. RMSEP (%) LEFT COLD 0.87 2.24 0.85 2.32 0.83 2.97 0.77 3.43 RIGHT WARM 0.91 1.83 0.91 1.99 0.79 3.08 0.65 4.02 RIGHT COLD 0.91 1.91 0.88 2.15 0.79 3.01 0.63 4.58 LEFT WARM 0.90 2.00 0.88 2.13 0.87 2.49 0.60 4.30 AVERAGE 0.90 2.00 0.88 2.15 0.82 2.89 0.66 4.08

(2.4) Effect of adding external variables

The results of Table 8 and FIG. 14 show that the performance of the model is improved when three external variables are included, although the performance increase is almost negligible. The slight increase in model performance may be due to the fact that already extracted features are related to external parameters. As a result, it was decided not to include three external factors in the model to simplify the model. In the predictive model, the influence of each external factor on the model was identified, and it was confirmed that the sex of the carcass was a more important characteristic than other age and weight. In fact, as can be seen in FIG. 12 , the gender characteristic (#3) was high enough to be similar to the weight of B10-90. However, since the function of the selected features showed overlapping characteristics with other features, the effect on the model was very negligible.

Selected Features Selected + External Features R ² cal. RMSEC R ² pred. RMSEP R ² cal. RMSEC R ² pred. RMSEP LEFT COLD 0.87 2.32 0.83 2.59 0.87 2.24 0.85 2.32 RIGHT WARM 0.92 1.67 0.89 1.99 0.91 1.83 0.91 1.99 RIGHT COLD 0.89 2.06 0.89 1.94 0.91 1.91 0.88 2.15 LEFT WARM 0.92 1.84 0.88 2.03 0.90 2.00 0.88 2.13 Average 0.90 1.97 0.87 2.14 0.90 2.00 0.88 2.15

(2.5) All Features vs Selected Features

Table 9 is a table comparing the model constructed using all the characteristics extracted from the temperature body likelihood model and the model constructed using the selected characteristics. As can be seen from the table, it can be confirmed that the performance of the MLR model built by selecting features is superior to the GPR-based model using all features. The GPR model for all features showed excellent results in calibration, but the actual prediction results were not good. In particular, the accuracy of calibration and validation showed a significant difference, and it was ^{confirmed that the temperature body R 2} increased by about 7% and RMSE by about 8.5%. On the other hand, in the model built with the selected features and a simple MLR model, the performance difference between the calibration set and the validation set was not large (R ² decreased by about 2% for the thermobody, and the RMSE decreased by about 1.6%). The 23 feature-based MLR model selected for LMP prediction ^{increased R 2} by 5% compared to the model using all features, and showed about 30% improvement in RMSE. This performance difference can also be clearly confirmed in the regression plot of FIG. 14 . In the residual plots, it was also observed that there was still a correlation between the residuals of the model using all features and the actual LMP values, but this correlation was not seen in the model in which the features were selected.

Also, similar to the results of LMP, a similar trend was observed for the whole and partial meat weight prediction items. As for the prediction accuracy, it was confirmed that the RMSE value was clearly improved in the model built based on the selected features rather than when all features were used. The reductions in RMSEP values achieved for the different prediction items are as follows. HCW (from 29.10 kg to 23.43 kg, using 8 variables), tenderloin (from 0.89 to 0.77 kg, using 23 variables); sirloin (from 4.96 to 3.10 kg, 25 variables); strip_stripe (from 1.32 to 1.31 kg, 28 variables); mokshim_chuck-roll (from 3.42 to 3.04 kg, 19 variables); foreleg_fore shank (from 2.79 to 2.24 kg, 18 variables); dumb_top round (from 2.43 to 1.53 kg, 19 variables); Seoldo_bottom round (from 3.44 to 2.61 kg, 21 variables); brisket (from 9.18 to 6.07 kg, 23 variables); sash_hind shank (from 1.68 to 1.25 kg, 22 variables); ribs (from 5.23 to 4.97 kg, 22 variables). These improvements can be seen in Table 9. The regression scatterplot obtained using the selected features can be confirmed through FIG. 15 .

Due to the characteristics of data (images), noise and related problems may be extracted during feature extraction, and such noise may cause feature overlap or collinearity during model construction. Model building that proceeds through collinearity or noisy datasets is difficult to find convergence. These factors add unnecessary complexity to the model, resulting in over-fitting the model. This explains why the calibration set results always show higher accuracy than the validation set. In conclusion, if the number of variables is reduced through SFS, unnecessary features and noise are removed, and the model becomes simpler and shows good performance.

All variables (1400, GPR) Selected variables (SFS, MLR) R ² c RMSEC* R ² p RMSEP* R ² c RMSEC* R ² p RMSEP* Select LMP (%) 0.89 2.45 0.84 2.95 0.92 1.67 0.89 1.99 23 HCW (kg) 0.97 15.75 (3.8%) 0.91 29.10 (7.1%) 0.94 24.03 (5.8%) 0.92 23.43 (5.6%) 8 Tenderloin wt. (kg) 0.90 0.60 (6.5%) 0.83 0.89 (9.6%) 0.91 0.51 (5.5%) 0.87 0.77 (8.4%) 23 Sirloin wt. (kg) 0.83 3.65 (9.2%) 0.80 4.96 (12.5%) 0.91 3.17 (8.0%) 0.88 3.10 (7.8%) 25 Striploin wt. (kg) 0.92 0.80 (6.6%) 0.82 1.32 (11.0%) 0.91 1.09 (9.1%) 0.86 1.31 (10.9%) 28 Chuck-roll wt. (kg) 0.89 2.78 (12.1%) 0.86 3.42 (14.9%) 0.90 3.16 (13.7%) 0.86 3.04 (13.2%) 19 Fore shank wt. (kg) 0.90 2.04 (6.3%) 0.85 2.79 (8.6%) 0.93 2.10 (6.5%) 0.88 2.24 (6.9%) 18 Top round wt. (kg) 0.88 1.69 (6.1%) 0.83 2.43 (8.7%) 0.92 1.60 (5.7%) 0.92 1.53 (5.5%) 20 Bot. round wt. (kg) 0.90 2.41 (5.5%) 0.85 3.44 (7.8%) 0.93 2.48 (5.6%) 0.91 2.61 (5.9%) 21 Brisket wt. (kg) 0.92 6.94 (8.0%) 0.88 9.18 (10.6%) 0.95 5.54 (6.4%) 0.93 6.07 (7.0%) 23 Hind shank wt. (kg) 0.91 1.10 (5.8%) 0.83 1.68 (8.8%) 0.92 1.14 (6.0%) 0.90 1.25 (6.6%) 22 Ribs wt. (kg) 0.89 4.17 (7.1%) 0.83 5.23 (9.1%) 0.89 4.92 (8.5%) 0.86 4.97 (8.6%) 22

Prediction results for the weights of LMP, HCW and 10 major versus split sites show that values can be estimated using less than 30 features (Table 9, FIGS. 14 and 15). The features extracted for the predictive parameters varied, but the B, D, and E features prevailed, constituting about 80% of the selected features. This implies that the feature extraction strategy of feature A and feature RP should be modified. In addition, more than 60% of the extracted features were extracted from image 1. This is because most of the features were extracted from image 1, and the number of features extracted from image 2 was about 20 times greater. 13 is a diagram of the features selected from the right conductor image 1 in the temperature body state.

In addition, it was confirmed that the selected characteristics were slightly deformed while the test on the model was repeated. This may be because there are many correlated variables that are easily interchangeable with each other. Therefore, in order to achieve higher numerical performance than the prediction accuracy confirmed in this experimental example, additional research on feature extraction is required. It was confirmed that the model prediction performance for the weight of large split meat was similar to the performance achieved in 3D-based BCC-3 (Jakob et al., 2019).

This experimental example demonstrated that it is possible to predict the meat quantity of cattle carcases through color images by combining multivariate statistics and image analysis. The performance of the model built through the characteristics selected through image analysis and MLR ^{was confirmed that R 2} was greater than 0.86 for all parameters. In particular, the prediction performance for LMP ^{showed high prediction accuracy with R 2} =0.89 and RMSE = 1.99%. This performance showed relatively good performance even when compared to most VIA systems such as VBS2000 and VIAScan. Therefore, it is judged that this system will be a landmark turning point for real-time carcase evaluation in slaughter lines.

In this experimental example, it was confirmed that the performance of the carcass rate prediction model was improved when additional external characteristics were added to the predictor variables. Additionally, it is possible to develop a predictive model using only one image during image analysis, but it is thought that using one or more images taken from different directions of the conductor helps to increase the model accuracy. In addition, we found that there is a possibility of showing much better model performance if further research on feature extraction is conducted.

Features, structures, effects, etc. exemplified in each of the above-described embodiments may be combined or modified for other embodiments by those of ordinary skill in the art to which the embodiments belong. Accordingly, the contents related to such combinations and modifications should be interpreted as being included in the scope of the present invention.

Claims (6)

an image collection step of obtaining images of conductors in various states using a color camera;
processing the image collected in the above step to extract features, the image processing step including a background removal step, a key point detection step and a feature extraction step; and
A model building step of constructing a predictive model using a multiple linear regression (MLR) method using the selected feature among the extracted features;
The image collection step is a step of obtaining two images including a conductor inward image and an image in the dorsal direction of a cold conductor or a temperature body on the side to which a tail is not attached among the left and right conductors,
The extracted features are the cross-section, center, and contour when the boundary box surrounding the conductor is subdivided into the same segment, the features extracted based on the characteristics of the center and the contour, using the mask image obtained through the background removal step. Includes features extracted based on a combination of features and key points,
The model building step uses at least one of a feature selection cleaning (FSC) algorithm and a sequential feature selection (SFS) algorithm as an algorithm for selecting a feature to be used for modeling among the extracted features. How to build a predictive model.
delete According to claim 1,
In the background removal step, the final mask image is obtained through a morphological image processing process for the image generated by removing the blue screen area and combining the threshold setting and edge detection after improving the conductor contrast through color statistics. A method of constructing a meat quantity prediction model of cattle carcass, characterized in that the step of removing the background.
According to claim 1,
In the key point detection step, a point that can be the starting point of classification for each part of the conductor is selected as a key point, a model conductor image is selected, a master image with key point labeling is specified, and when a new image is collected, the master image and A method of constructing a meat quantity prediction model of cattle carcass, characterized in that the step of mapping the key point position to a new image by matching.
delete delete

Images