Introduction
The Agricultural Outlook of 2025 to 2034 from the Organisation for Economic Co-operation and Development and the Food and Agriculture Organization of the United Nations (OECD/FAO; 2025) projects a 6% increase in global per capita consumption of animal-source foods by 2034 (beef, pork, poultry, fish, dairy, and other animal products) and is most pronounced in lower middle-income countries, where it is expected to rise by 24%, far outpacing the global average. The swine industry in Canada holds a 4.17% share of the global market, ranking as major producer and exporter (Mali, 2026). Therefore, it is imperative to satisfy this demand, which is a worldwide challenge for future food security. In addition, since 2020, the Canadian swine industry has faced unprecedented challenges due to COVID-19, including difficulty in marketing finished pigs and labor shortages (Hein, 2020). Thus, minimizing input waste and costs is a critical strategy for improving farm efficiency and viability, enabling farmers to exercise greater control over their resources. Achieving optimal input efficiency is vital for all producers, as higher technical efficiency reduces resource use, lowers production costs, and increases profitability, which serves as a key motivator for adopting innovative practices (Galanopoulos et al., 2006).
Pork producers select pigs for slaughter based on live pig conformation and weight, which is labor intensive and stressful for the pigs. However, most hog markets are based on pork carcass merit, determined postmortem by lean-grade index. The lean-grade index (80–116) combines hot-carcass weight with predicted lean percentage (Pomar and Marcoux, 2003). This index is used in payment grids to reward producers for desirable pork carcasses and apply discounts for undesirable ones, such as those that are too light, heavy, lean, or fatty (Carr et al., 1997; Zhou and Bohrer, 2019). Consequently, many pigs do not meet packing plant specifications, resulting in value loss for hog carcass nonconformities of averaged $8.08 per carcass (e.g., inconsistent live weight, thin bellies, too fat carcass and wholesale cut weights) (Meisinger, 2003). However, in the last pork quality audit (Boler, 2017), packers were less willing to share information about their system, so recent value loss is unknown. Leanness is assessed through invasive techniques on the carcass, which cannot be used on farms to select market animals. Therefore, a noninvasive real-time monitoring system for growth rates and in-vivo carcass composition could enable producers to make real-time management adjustments (e.g., diet) as pigs grow, improving resource use. The proposed technology can promote sustainable intensification, enhance animal welfare (e.g., reducing animal handling stress and injuries), and alleviate the shortage of specialized labor.
Recent efforts have shown the potential of 2-dimensional, 3-dimensional (3D), multispectral, and hyperspectral imaging analysis in chicken, pig, beef, bison, and dairy cattle to estimate health, behavior, milking traits, body, carcass, and meat characteristics (Sun, 2016; Benjamin and Yik, 2019; Miller et al., 2019; Chaudhry et al., 2021; Hasan et al., 2022). However, those technologies require adequate lighting, and a contrasting background (dark floor or wall), avoidance of visual interference (facilities, equipment, and operators), and high device, operational, and maintenance costs may also create challenges for producers (Martins et al., 2020; Carrara et al., 2024; Nisbet et al., 2024).
On the other hand, the thermal infrared image has been used in the livestock sector to identify metabolically efficient animals (Schaefer et al., 2023), monitor animal health, physiology, and welfare (Hurnik et al., 1985; Schaefer et al., 2004; Schaefer et al., 2007; Schaefer et al., 2025), predict beef meat quality defects (Cuthbertson et al. 2020), monitor beef carcass cooling (Hite et al., 2020), and predict lamb cuts’ weights (de Souza et al., 2025). A thermal camera assesses skin surface temperature and is an indirect method for assessing body-fat composition (by decreasing skin surface temperature), but it does not provide a direct measure of fat-to-muscle ratio (Salamunesa et al., 2017). However, most of these studies used manual image preprocessing and feature extraction methods (e.g., selecting the region of interest in each image to obtain the minimum, median, and maximum temperatures of the area), which are often time consuming. Thus, applying a modern artificial intelligence technique can reduce image processing time, capture spatial hierarchies and local features (i.e., thermal patterns), and predict carcass parameters from thermal images.
Therefore, this study aims to predict in-vivo carcass traits and carcass leanness levels using thermal imaging, computer vision models (CVM) to classify growing pigs into leanness classes (higher and lower lean-grade indices defined based on carcass weight and estimated percent lean yield for a pork carcass), and to compare the lean-yield classification obtained by dual-energy x-ray absorptiometry (DEXA) vs. electronic grading probe (Destron).
Materials and Methods
The University of Manitoba Animal Care Committee (F20-026) approved the experimental procedures, ensuring they conform to Canadian Council on Animal Care guidelines (Olfert et al., 2009).
Animal and Carcass Management
A total of 243 hundred finishing pigs (barrow and gilt; Large White × Landrace sows × Duroc boars; Genesus Genetic Technology, MB, Canada) were evaluated in a swine unit located in Alberta, Canada (Lacombe Research Centre). The pigs were fed a corn-wheat commercial finisher diet, formulated to meet the nutrient requirements by the National Research Council (NRC, 1998). Throughout the experiment, animals were provided with unrestricted access to feed and water. When animals reached the endpoint of 122 kg ± 7, images of the pigs were captured, and the pigs were slaughtered.
Image Acquisition
The pigs were calmly moved through the chute using rattle paddles in a single file toward the imaging area, where the scale and camera were located. Once the pig was placed on the scale, the image of each pig was captured from the dorsal view. One dorsal image (640 × 512 pixels; Figure 1) per pig was captured from all pigs at a focal distance of approximately 2 m with a thermal camera (between 5–15 μm; longwave infrared spectrum, capture speed of 30 Hz; FLIR A 65 [FLIR System, Inc., Wilsonville, Oregon]) (Taylor et al., 2023). A total of 243 images were captured. The setup for image acquisition (Figure 2) consisted of an overhead camera, fluorescent lights (Philips 28190-7 25W, 3500 k, CRI 85) at 61 l ×, and a nonreflective, uniform background. Each pig was assigned a unique radio frequency identification tag to track the images and their corresponding data. The temperature and humidity in the scale room were monitored (15–20°C and 25–30%, respectively) and used to calibrate the camera at every image capture session.
Dorsal images were captured with a multispectral camera (500 X 600 pixel) infrared detection system (between 5–15 μM; capture speed of 30 Hz; FLIR A65 Comp. Boston, MA, USA). Source: Agriculture and Agri-Food Canada.
Set up for image acquisition.
Slaughter and Carcass Fabrication
Three days after the images were taken, pigs were slaughtered. After dressing, carcasses were split into 2 halves and weighed, and a Destron electronic probe Viewtrak PG-309 (Viewtrak Technologies Inc., Markham, ON, Canada) was applied to the left carcass sides to determine fat depth, lean depth, and leanness. The Destron probe was placed perpendicularly into the carcass between the third and fourth last rib, approximately 7 cm off midline, to measure the fat depth and lean tissue (Barducci et al., 2020). Carcass sides were then moved into the cooler with temperature set to 2°C. After a 24-h postmortem period, chilled carcasses were weighed and then left carcass sides were divided into primal cuts (picnic shoulder, butt, ham, loin, and belly; CPI, 1995). Primals were scanned by DEXA to measure leanness and total yield and coproduct components, which were calculated as percentages of the cold-carcass weight as described by Soladoye et al. (2016).
Data Preprocessing
Data collected were analyzed using R (R Core Team, 2023). Descriptive statistics of central tendency and dispersion were determined for the studied variables. Additionally, considering DEXA leanness as the gold standard for its high accuracy (R2 > 0.90) with the manual dissection method (Soladoye et al., 2016), Destron leanness was contrasted with DEXA leanness through a linear regression model. It was used: Yj = β0 + β1 Destronj + ɛj, in which Yj represented the lean yield measured by DEXA (%), β0 was the intercept, β1 was the linear regression coefficient, Destronj was the coefficient associated with the lean yield measured by Destron (%), and ɛj was the residual error ∼ N (0, σ2). The coefficient of determination (R2) and P value were used to evaluate the model. The statistical significance level was set at a α value of less than 0.05. A scatter plot was used to visualize the relationship between the 2 variables.
The lean-grade index was determined for the carcasses. Lean-grade index (80–116 scores) is a combination of the hot-carcass weight and the lean percentage, where carcasses were classified into higher (>109 scores, where carcasses had between 57.7–64.2% of leanness and hot-carcass weight ranging from 80–105 kg) and low lean-grade indexes (<109 scores) (Pomar and Marcoux, 2003). This was done using the measure of carcass leanness obtained from DEXA and Destron. The obtained classes from both methodologies were compared as well as used as labels for training the models. χ2 analysis was used to test differences in frequencies of lean-grade index classes obtained by DEXA and Destron.
A total of 238 images were available with complete carcass composition data. Using an open-source pretrained segmentation model (Ravi et al., 2024), the images were segmented to remove the background and maintain only the pigs (i.e., the region of interest) as suggested by Zheng et al. (2014). As a result, the images varied in size due to the segmentation process, and a padding technique was implemented by adding zeros (i.e., empty pixels) based on the largest image to ensure all images were the same size to train the models. The final images were 296 × 610 pixels and had 1 channel (i.e., black and white color).
The image dataset was divided into 3 sets: training for model development, validation for hyperparameter tuning and evaluation during training, and testing to assess the model’s capacity to generalize to animals not seen during model development. The split was done using the ratio of 70%, 15%, and 15%, respectively, for the train, validation, and test sets. The split was stratified based on the distribution of the outcome variables.
Data augmentation (Figure 3) was implemented in the training set as a strategy to increase the number of observations (165–990) available for model training, as suggested by Hodnett and Wiley (2018). The following augmentation procedures were implemented: a horizontal flip of the images at 180° was applied; shifting 1: shifted the image 50 pixels to the x-axis and 50 pixels to the y-axis; shifting 2: shifted the image −80 pixels to the x-axis and −80 pixels to the y-axis; and blurring: a Gaussian blur with a kernel size of 1 × 15 pixels and a standard deviation of 0.5 for the blur intensity. Noise was added by sampling from a Gaussian distribution (mean = 0, SD = 1) and scaling it by 0.80 to control its magnitude relative to the original images. After adding noise, pixel values were normalized back to the original range of 0 (i.e., fully black) to 1 (i.e., fully white).
Various data augmentation strategies were implemented: (A) shows the original image; (B) the image is flipped over by 180° horizontally; (C) blur dimensions were added to the image; (D) the image was shifted by 50 pixels; (E) the image was shifted by −80 pixels; and (F) noise from a Gaussian distribution (mean = 0; SD = 1) was added to the image.
Model Training
CVM based on convolutional neural networks (CNN) were trained for regression (prediction of individual variables) and classification analysis (high or low lean-grade index). The outcome variables fat depth, lean depth, total lean, and total fat were used in the regression models. For the classification models, carcass classes determined based on either DEXA or Destron were used as outcome variables. For both analyses, the thermal images were used as input to train the CNN models. In Table 1, the number of images per class is shown before and after data augmentation and data splits.
The number of images per data split as having a low and a high lean-grade index
| Carcass Score1 | Train (Before Augmentation) | Train (After Augmentation) | Validation | Test |
|---|---|---|---|---|
| DEXA | ||||
| Low lean-grade index | 96 | 576 | 20 | 21 |
| High lean-grade index | 69 | 414 | 16 | 16 |
| Destron | ||||
| Low lean-grade index | 71 | 426 | 16 | 16 |
| High lean-grade index | 94 | 564 | 21 | 20 |
Abbreviation: DEXA, dual-energy x-ray absorptiometry.
The percentage of higher and lower grade indexes obtained based on carcass leanness and hot-carcass weight. Leanness was measured separately using DEXA and Destron methods.
We used a Bayesian optimization procedure to select the hyperparameters for the models (Eriksson and Poloczek, 2021; Nguyen, 2019). The hyperparameters evaluated were as follows: the number of convolutional layers (1–4), the number of filters in the convolutional layers (32, 64, or 128), the kernel size (3, 5, or 7), the number of dense layers (1–3), and the number of units in the dense layers (64 or 128). Early stop and learning rate reduction on the plateau were implemented to avoid overfitting and improve model learning, respectively. The maximum number of epochs during hyperparameter optimization was set to 250, but the training was stopped if the validation loss did not decrease after 15 epochs. The learning rate was decreased by a factor of 0.1 until a minimum rate of 1 × 10−5 was reached if the training loss did not decrease after 10 epochs. Once the best hyperparameter combination was identified, the final models were also trained for a maximum epoch number of 250, along with early stop and learning rate reduction on the plateau. Final model training was repeated 5 times to account for the random initializations of weights at the beginning of the training process of the CNN.
Model Evaluation
For the regression analysis, models were evaluated based on the root mean squared error (RMSE), mean absolute error (MAE), and the out-of-sample coefficient of determination (osR2) (Hawinkel et al., 2024). For RMSE and MAE, lower values are desired as they indicate how close the predicted value is to the actual value. For osR2, which ranges from −∞ to 1, values closer to 1 indicate a better predictive capacity. For the classification analysis, the models were evaluated using precision, accuracy, recall, and F1 score, all of which range from 0 to 1. The closer to 1, the better the model’s predictive accuracy. A confusion matrix for the test set was also produced to provide a detailed breakdown of the models’ performance.
Results
Descriptive Statistics of Carcass Traits and Composition Variables
Table 2 presents the descriptive statistics of carcass traits and composition variables observed in this study. The data indicate a broad range of carcass fatness traits (fat depth and total fat percentage), which corresponds with moderate variation (coefficient of variation [CV] >10%). Conversely, lower variation was observed (CV < 10%) for slaughter and carcass weights, lean depth, and total lean percentage.
Descriptive statistics of carcass traits and composition variables (N = 238)
| Variable | Mean | STD | CV | Min | Median | Max |
|---|---|---|---|---|---|---|
| Slaughter wt, kg | 121.6 | 6.55 | 5.39 | 102.5 | 122.0 | 145.0 |
| Hot-carcass wt, kg | 102.1 | 5.74 | 5.62 | 88.6 | 102.1 | 122.0 |
| Fat depth, mm1 | 23.0 | 5.21 | 22.65 | 12.0 | 22.4 | 45.3 |
| Lean depth, mm1 | 63.7 | 5.81 | 9.12 | 40.5 | 64.1 | 78.3 |
| Total lean, %2 | 57.7 | 3.50 | 6.07 | 47.2 | 58.1 | 66.0 |
| Total fat, %2 | 32.6 | 3.74 | 11.47 | 23.2 | 32.5 | 44.0 |
Abbreviations: CV, coefficient of variation; DEXA, dual-energy x-ray absorptiometry; Max, maximum; Min, minimum; STD, standard deviation.
Measured with Destron electronic probe.
Total lean muscle and fat measured using DEXA.
Contrasting Dual-Energy X-Ray Absorptiometry and Destron Based on Lean Yields and Lean-Grade Index
The scatter plot in Figure 4 shows the relationship between the DEXA and Destron lean yields. A moderate relationship (P < .001; R2 = 0.64) was observed between lean-yield estimates from DEXA (as standard method) and Destron, indicating that Destron was less accurate. Also, it was denoted in the distribution of the carcasses segregated in low and a high lean-grade index (Table 3; P < .01) that a high percentage of animals were considered as high lean score based on Destron. This percentage might indicate that Destron overestimates carcass leanness and could not be used as an alternative predictor to train the models.
Contrasting lean yield obtained by dual-energy x-ray absorptiometry vs. lean yield from Destron. Abbreviation: DEXA, dual-energy x-ray absorptiometry.
The distribution of the carcass classes1 as having a low and a high lean-grade index (N = 238)
| Carcass Score2 | N | % |
|---|---|---|
| DEXA | ||
| Low lean-grade index | 137 | 57.6 |
| High lean-grade index | 101 | 42.4 |
| Destron | ||
| Low lean-grade index | 103 | 43.3 |
| High lean-grade index | 135 | 56.7 |
Abbreviation: DEXA, dual-energy x-ray absorptiometry.
χ2 analysis indicated that the distribution of lean-grade index classes differed between DEXA and Destron (P < .01).
The percentage of higher and lower grade indexes obtained based on carcass leanness and hot-carcass weight. Leanness was measured separately using DEXA and Destron methods.
Predicting Individual Traits
The CVM CNN models trained for the regression analysis to predict individual carcass leanness- and fatness-related traits performed poorly (oSR2 < 0.30; Table 4). Although lean depth performed moderately during the training stage (oSR2 = 0.62), its performance decreased to low in the validation and testing stages (oSR2 < 0.30).
The performance of convolutional neural network models in predicting fat and lean depth and total lean and fat across the dataset
| Variable | Dataset | RMSE | MAE | osR2 |
|---|---|---|---|---|
| Fat depth, mm1 | Training | 5.02 ± 0.03 | 3.97 ± 0.04 | 0.12 ± 0.01 |
| Validation | 4.67 ± 0.02 | 3.83 ± 0.03 | 0.01 ± 0.01 | |
| Testing | 4.93 ± 0.08 | 3.86 ± 0.11 | 0.04 ± 0.03 | |
| Lean depth, mm1 | Training | 3.34 ± 0.67 | 2.59 ± 0.52 | 0.62 ± 0.14 |
| Validation | 6.22 ± 0.10 | 4.96 ± 0.16 | 0.27 ± 0.02 | |
| Testing | 5.77 ± 0.11 | 4.65 ± 0.19 | −0.14 ± 0.04 | |
| Total fat, %2 | Training | 2.93 ± 0.47 | 2.33 ± 0.40 | 0.38 ± 0.19 |
| Validation | 3.44 ± 0.05 | 2.72 ± 0.06 | −0.003 ± 0.03 | |
| Testing | 4.19 ± 0.23 | 3.19 ± 0.21 | −0.20 ± 0.13 | |
| Total lean, %2 | Training | 3.16 ± 0.09 | 2.45 ± 0.06 | 0.15 ± 0.06 |
| Validation | 2.95 ± 0.17 | 2.23 ± 0.03 | 0.28 ± 0.02 | |
| Testing | 3.55 ± 0.04 | 2.81 ± 0.03 | 0.09 ± 0.02 |
Abbreviations: DEXA, dual-energy x-ray absorptiometry; MAE, mean absolute error; osR2, out-of-sample coefficient of determination; RMSE, root mean squared error.
Measured with the Destron electronic.
Total lean muscle and fat measured by using DEXA.
Classification Analysis
The performance of the CVM CNN models trained for the classification analysis to segregate between a high and a low lean-grade index (>109 scores) using DEXA and Destron lean methods are shown in Table 5. CVM using DEXA had moderate performance, with an average F1 score of 0.73 on the testing set. However, the F1 score for the lean-grade index based on the Destron leanness assessment was low (F1 score = 0.38) in the testing set, showing that the model misclassifies a significant number of images. In addition, overfitting was also observed when training the CNN model based on the Destron leanness assessment, as indicated by a large difference between the model performance on the training set compared to the validation and testing sets, despite the different strategies employed to avoid overfitting (Table 5). Additionally, the confusion matrix (Table 6) showed that the DEXA model demonstrated a bias toward predicting low lean-grade index, while the Destron model showed no clear predictive pattern, consistent with the observed overfitting and low F1 score.
Model performance in classifying carcasses according to lean-grade index using 2 methods of lean determination
| Variable | Set | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|---|
| DEXA lean-grade index | Training | 0.69 ± 0.07 | 0.67 ± 0.06 | 0.93 ± 0.05 | 0.78 ± 0.03 |
| Validation | 0.63 ± 0.05 | 0.63 ± 0.06 | 0.86 ± 0.13 | 0.72 ± 0.02 | |
| Test | 0.63 ± 0.04 | 0.62 ± 0.04 | 0.90 ± 0.11 | 0.73 ± 0.02 | |
| Destron lean-grade index | Training | 0.92 ± 0.01 | 0.99 ± 0.01 | 0.82 ± 0.01 | 0.89 ± 0.01 |
| Validation | 0.64 ± 0.03 | 0.69 ± 0.08 | 0.30 ± 0.10 | 0.41 ± 0.11 | |
| Test | 0.52 ± 0.02 | 0.45 ± 0.04 | 0.32 ± 0.05 | 0.38 ± 0.05 |
Abbreviation: DEXA, dual-energy x-ray absorptiometry.
10–116 scores, segregating higher lean-grade indexes (>109 scores; carcasses between 57.7–64.2% of leanness and 80–105 kg of hot-carcass weight).
Confusion matrix of dual-energy x-ray absorptiometry and Destron on test set
| Observed | |||
|---|---|---|---|
| DEXA | Low Lean-Grade Index | High Lean-Grade Index | |
| Predicted | Low lean-grade index | 20 | 12 |
| High lean-grade index | 1 | 4 | |
| Destron | |||
| Low lean-grade index | 5 | 7 | |
| High lean-grade index | 11 | 13 | |
Abbreviation: DEXA, dual-energy x-ray absorptiometry.
10–116 scores, segregating higher lean-grade indexes (>109 scores; carcasses between 57.7–64.2% of leanness and 80–105 kg of hot-carcass weight).
Values are averaged over the 5 models that were trained.
Discussion
Prediction of Individual Traits
The models performed poorly (R2 ≤ 0.09) in predicting individual traits (i.e., fat depth, total lean). To the best of our knowledge, no previous studies have specifically used thermal imaging technology and CVM to evaluate in-vivo body or carcass traits and composition in livestock. However, studies in humans evaluating fat distribution have shown promising results. Bauer et al. (2020) applied infrared thermography combined with artificial intelligence (artificial neural network) to detect both early and advanced cellulite stages (subcutaneous fat deposition, connective tissue, and fluid buildup, especially common on thighs and buttocks in women), where a combination of histogram of oriented gradients and artificial neural network enables determination of all stages of cellulite with an average accuracy higher than 80%. On the other hand, Snekhalatha et al. (2021) applied deep learning techniques to identify obesity using thermal images, evaluating both pretrained and custom CNN models, and notably, their CNN Custom-2 model reached a classification accuracy of 92% when differentiating between obese and normal subjects. Similarly, comparable accuracy (83%) can be obtained by applying lightweight CNN models (DenseNet201) for obesity early detection using thermal images (Leo et al., 2024).
On the other hand, in livestock, the use of thermal imaging to evaluate body or carcass composition is scarce. A study (de Souza et al., 2025) developed predictive models for the weights of lamb lean primal cuts using morphometric measurements (external carcass length, thoracic depth, leg length, and leg width) extracted from carcass surface thermal images and found high accuracy and precision combining morphometric measurements to predict lean weight from a high-value cut, topside (R2 = 0.81); however, the other cuts (flank steak, rump cap) were predicted moderately to low (R2 = 0.69–0.41). Similar results (accuracy 72–85%) can be obtained using extracted morphometric measurements from RGB (red-green-blue pixel matrixvalues) and video images of sheep carcasses to predict cuts such as shoulder, rib, and loin (Brady et al., 2003; Rius-Vilarrasa et al., 2009; Ngo et al., 2016; Batista et al., 2021).
Comparing other predicting technologies, Fernandes et al. (2020) developed a computer vision system to predict muscle depth, and backfat thickness measured by ultrasound (Aloka SSD 500, 3.5-MHz, 12-cm linear probe) using top-view 3D images (Microsoft Kinect V2) of finishing pigs by obtaining a total of 12 000 images from 557 finishing pigs. This approach was likely aimed at capturing a comprehensive dataset for each animal, which could potentially improve model accuracy, and deep learning models achieved R2 values of 0.50 for muscle depth and 0.45 for backfat (key indicators of lean muscle mass). Masoumi et al. (2021) found moderate predictability for lean and fat weight (R2 = 0.77 and 0.73, respectively) using digital imaging developed from full 3D models of half carcass sides of pigs, obtained by scanning each half carcass side using the Go!SCAN 3D TM (Model 50, Creaform, Levis, Quebec, Canada). Lohumi et al. (2018) predicted lean meat yield in commercial pork carcasses with moderate accuracy (R2 = 0.77) using a vision-based video image analyzer (VCS2000) and extracting morphometric measurements from a smaller dataset, which may have influenced model performance. A similar approach (video and extracting morphometric measurements) was used by Miller et al. (2019) and demonstrated that 3D imaging coupled with machine learning analytics (artificial neural networks) can be used to predict liveweight (R2 = 0.70), cold-carcass weight (R2 = 0.88), and saleable meat yield (R2 = 0.72) in live beef cattle.
Carcass Lean-Grade Index Classification
On the other hand, models trained for classification analysis showed improved performance, particularly when using DEXA. Studies evaluating the robustness of classification algorithms have been identified. Pei et al. (2023) examined the robustness of machine learning algorithms to predict the characteristics of optical micrographs. They found that while the classification model achieved a 90.79% accuracy, the regression model obtained an R2 of 0.21. Furthermore, in a comparative study of classification and regression algorithms for modeling student academic performance, Strecht et al. (2015) found that classification algorithms successfully identified useful patterns, while regression models failed to outperform a simple baseline. The classification model demonstrated greater robustness, with a lower standard deviation (0.17 compared to 0.20). In contrast, all regression algorithms produced an error of approximately 5, which was considered very high on the 0 to 20 scale used in the study.
In the current study, the classification model demonstrated moderate performance in classifying carcasses into high and low lean-grade groups based on DEXA-determined lean-grade indexes (F1 score; 0.73 vs. 0.38). The DEXA device can assess the entire body to capture detailed information on tissue distribution and density, enabling accurate estimates of body composition. This may be why the classification model predictions from DEXA measurements performed more accurately than those based on the Destron, which measures backfat thickness and muscle depth at 1 location on the carcass using light reflectance technology and then calculating the leanness of the whole carcass (Kipper et al., 2019). The model based on Destron variables demonstrated some level of overfitting despite the regularization procedures implemented: it fit the training set well (F1 = 0.89) but performed poorly on the test set (F1 = 0.38).
Publications comparing Destron and DEXA predictions do not exist to the authors’ knowledge. However, under Canadian swine production conditions, Destron has been compared with carcass dissection (Usborne et al., 1987, Pomar et al., 2001), other optical grading probes such as Henessy grading probe (Pomar and Marcoux, 2003) and advanced automated ultrasonic scanner (AutoFom III; Bohrer et al., 2023), which have remarked certain deficiencies. Previous studies (Usborne et al., 1987; Pomar et al., 2001) have identified low accuracy (R2 < 0.56) of the Destron in predicting lean yield determined by carcass dissection. Also, Pomar and Marcoux (2003) found that the Destron method consistently predicts lower lean-yield and carcass-index values than the Hennessy method, and these differences become more pronounced as carcass leanness increases. Bohrer et al. (2023) updated the Destron lean-yield equation and compared it with the existing Destron lean-yield equation on precision and accuracy, finding that both equations had similar predictive precision (R2 = 0.75). However, in terms of accuracy, the existing equation estimated lean yield correctly 8.1% of the time, while the updated equation did so 47.7% of the time. For that reason, we considered DEXA leanness the gold standard for the current study, given its high accuracy (R2 > 0.90) relative to the manual dissection method (Soladoye et al., 2016).
Limitations
This substantial decrease in osR2 indicated a clear pattern of overfitting, where the model has memorized the training data, including its noise, but fails to generalize effectively to new data in the test set (Hawkins, 2004). Furthermore, the model’s predictive accuracy for total fat (%) and total lean (%) declined from training to testing, indicating overfitting and poor generalization. Overfitting happens when a model is too complex for the data. For instance, using a neural network for linear data adds unnecessary complexity. This extra flexibility does not improve performance and can lead to poorer predictions compared to a simpler model (Hawkins, 2004). Nonetheless, CNN models are well-suited for image data because they capture spatial hierarchies and local features effectively. In our study, the observed overfitting indicates that further measures are needed, such as increasing the number of augmentation strategies and adding regularization layers (L1, L2, and dropout) to avoid overfitting during training and improve their generalization.
The data size could explain the moderate performance in classifying the lean-grade index by DEXA. Gygi et al. (2023) noted that increasing the training dataset size often enhances prediction accuracy, adding that diverse data help models generalize better to less-represented scenarios. Furthermore, Chen et al. (2022) noted that limited datasets can lead to overfitting and emphasized the importance of developing models that identify consistent relationships, which is more effectively achieved by incorporating diverse populations into the training data. Also, Chen et al. (2022) revealed that overfitting in CNN arises from excessively large models, which are influenced by the size and number of convolutional kernels, components learned during training and are responsible for detecting patterns such as edges or textures. Larger kernels capture more information but increase the number of parameters, complexity, and processing time. Smaller kernels reduce the visual range but enhance speed. Because the kernel number (i.e., filters) and size were optimized in our study, the dataset size was likely a major factor explaining the observed model performance.
To evaluate the most suitable model design for our domain-specific dataset, we developed a custom CNN architecture optimized via Bayesian search rather than using pretrained ImageNet networks. Although pretrained models can be advantageous due to their broad applicability and transfer learning benefits, their convolutional filters are tuned to images that may not directly match the structural properties of our thermal imagery. Conversely, Bayesian optimization enabled us to systematically explore architectural parameters, including depth, kernel size, and filter count, thereby identifying a configuration tailored to the resolution of our images. This approach reduced model complexity and computational costs, providing a more resource-efficient alternative that better suits our specialized data and can more easily meet edge inference requirements, which is likely the scenario in which such models would be deployed to support management decisions.
In the current study, images were captured from the dorsal view of the pigs due to logistical constraints (e.g., equipment obstructions); however, missing lateral and hind-view images might not affect the prediction accuracy of body or carcass composition obtained in our study. Afonso et al. (2024) used video image analysis (VIA) to evaluate the composition of light lamb carcasses from images captured from the lateral and dorsal carcass sides, where dimensions such as lengths, widths, angles, areas, and perimeters were extracted, and K-Folds stepwise multiple regression analyses were employed to construct prediction models for carcass-tissue weights (including muscle, subcutaneous fat, intermuscular fat, and bone) and their respective percentages. The author reported that only the dorsal view was able to predict carcass-tissue weight (R2 = 0.94; lean, fat and muscle) rather than lateral view (R2 = 0.88); however, the prediction of carcass-tissue percentage from both dorsal and lateral views were from poor to moderate predictors (R2 > 0.70) even when they were combined. Also, Araújo et al. (2022) applied VIA and extracted biometric measurements from dorsal and lateral view images of live lambs. They showed higher to moderate accuracy in estimating hot- and cold-carcass weight and commercial-cuts weight using dorsal view data rather than lateral view data. The author explained that the dorsal region represents a large part of the carcass, with a high deposition of musculature and adipose tissue and tends to suffer less distortion when measured along the animal than when measured in side view, thus being less subject to animal positioning.
Conclusions
The present study evaluated the potential application of thermal imaging technology in conjunction with CNN to predict carcass lean and fat content in growing-finishing pigs. The regression CNN models performed poorly when predicting individual traits. However, a moderate performance was found when a classification CNN model was trained to segregate animals based on the lean-grade index obtained using the DEXA leanness. In addition, Destron was less accurate in determining carcass leanness than DEXA (gold standard), which might explain the poor performance of the classification CNN models in segregating animals based on the lean-grade index obtained using the Destron.
In Canada, Destron is the common instrument used in the industry to predict lean yield, which was approved and used since 1994; however, our findings showed that the instrument could not match with the level of leanness from the pork carcass cutout. Due to constant advances in genetics, nutrition, and management applied in swine production since 1994, future studies should be carried out to recalibrate the predicted lean yield from Destron with the level of leanness from the current pork carcass population.
Finally, thermal imaging technology could support farmers in marketing their hogs based on the best grid grade (carcasses with >109 scores) and increase the percentage of pigs meeting packing plant specifications and receiving premium prices. However, a larger and more diverse dataset is needed to improve the predictive ability of the models, particularly when predicting individual traits.
Conflict of Interest
Animal Inframetrics and Alpha-Phenomics participated in funding acquisition, conceptualization, investigation, and writing (reviewing and editing) and could be considered as having potential conflicts of interest due to their partial financial contributions to this work. However, the research analysis was conducted in the absence of any commercial or financial relationships between the co-authors and the sponsors.
Acknowledgment
The authors are grateful to Results Driven Agriculture Research and Alberta Innovates (Grant #212201006), Animal Inframetrics (in-kind contribution), and Alpha-Phenomics (in-kind contribution) for providing the funding to conduct this research.
Author Contribution
V. Ndams: investigation, formal analysis, and writing (original draft); G.M. Dallago: conceptualization, methodology, investigation, formal analysis, and writing (reviewing and editing); M. Nyachoti, C. Yang, C. Narvaez-Bravo, C. Grant, and V. A. M. Weber: methodology and writing (reviewing and editing); M. Juárez, O. López-Campos, and N. Prieto: methodology, investigation, and writing (reviewing and editing); A. Schaeffer and G. Plastow: funding acquisition, conceptualization, methodology, investigation, and writing (reviewing and editing); and A. Rodas-González: funding acquisition, conceptualization, methodology, project administration, supervision, investigation, formal analysis, and writing (reviewing and editing).
Literature Cited
Afonso, J. J., M. Almeida, A. C. Batista, C. Guedes, A. Teixeira, S. Silva, and V. Santos. 2024. Using image analysis technique for predicting light lamb carcass composition. Animals. 14:1593. doi: https://doi.org/10.3390/ani14111593.
Araújo, J. C., H. A. S. Santos, E. S. C. Ribeiro, A. C. C. Trindade, M. A. Sousa, M. P. M. Nunes, A. C. S. Lima, L. C. C. Daher, and A. G. M. Silva. 2022. Use of in vivo video image analysis as a substitute for manual biometric measurements on the prediction of qualitative and quantitative carcass characteristics of hair sheep lambs. Small Ruminant Res. 215:106779. doi: https://doi.org/10.1016/j.smallrumres.2022.106779.
Barducci, R. S., Z. Y. Zhou, L. Wormsbecher, C. Roehrig, D. Tulpan, and B. M. Bohrer. 2020. The relationship of pork carcass weight and leanness parameters in the Ontario commercial pork industry. Translational Animal Science. 4:331–338. doi: https://doi.org/10.1093/tas/txz169.
Batista, A. C., V. Santos, J. Afonso, C. Guedes, J. Azevedo, A. Teixeira, and S. Silva. 2021. Evaluation of an image analysis approach to predicting primal cuts and lean in light lamb carcasses. Animals. 11:1368. doi: https://doi.org/10.3390/ani11051368.
Bauer, J., Md. N. Hoq, J. Mulcahy, S. A. M. Tofail, F. Gulshan, C. Silien, H. Podbielska, and Md. M. Akbar. 2020. Implementation of artificial intelligence and non-contact infrared thermography for prediction and personalized automatic identification of different stages of cellulite. The EPMA Journal. 11:17–29. doi: https://doi.org/10.1007/s13167-020-00199-x.
Benjamin, M., and S. Yik. 2019. Precision livestock farming in swine welfare: a review for swine practitioners. Animals. 9:133. doi: https://doi.org/10.3390/ani9040133.
Bohrer, B. M., Y. Wang, J. B. Dorleku, C. P. Campbell, and I. B. Mandell. 2023. An update of the predicted lean yield equation for the Destron PG-100 optical grading probe. J. Anim. Sci. 101:skad199. doi: https://doi.org/10.1093/jas/skad199.
Boler, D. 2017. Pork supply chain audit. National Pork Board. https://www.porkcheckoff.org/wp-content/uploads/2021/02/16-115-BOLER-final-report.pdf. (Accessed 5 July 2021).https://www.porkcheckoff.org/wp-content/uploads/2021/02/16-115-BOLER-final-report.pdf
Brady, A. S., K. E. Belk, S. B. LeValley, N. L. Dalsted, J. A. Scanga, J. D. Tatum, and G. C. Smith. 2003. An evaluation of the lamb vision system as a predictor of lamb carcass red meat yield percentage. J. Anim. Sci. 81:1488–1498. doi: https://doi.org/10.2527/2003.8161488x.
Carr, T. R., R. G. Kauffman, D. L. Meeker, and D. J. Meisinger. 1997. Factors reducing pork value. In: Pork Industry Handbook, PIH-135. Purdue University Cooperative Extension Service, West Lafayette, IN. pp. 1–4.
Carrara, E. R., P. d. C. S. Oliveira, L. C. d. C. M. Dias, W. G. da Costa, A. R. Conceição, P. H. S. Braga, M. L. Chizzotti, R. Veroneze, and E. B. Schultz. 2024. Comparison of supervised machine learning and variable selection methods for body weight prediction of growth pigs using image processing data. Rev. Bras. Zootecn. 53:e20240001. doi: https://doi.org/10.37496/rbz5320240001.
Chaudhry, M. M. A., Md. M. Hasan, C. Erkinbaev, J. Paliwal, S. Suman, and A. Rodas-Gonzalez. 2021. Bison muscle discrimination and color stability prediction using near-infrared hyperspectral imaging. Biosyst. Eng. 209:1–13. doi: https://doi.org/10.1016/j.biosystemseng.2021.06.010.
Chen, H., L. Geng, H. Zhao, C. Zhao, and A. Liu. 2022. Image recognition algorithm based on artificial intelligence. Neural Comput. Appl. 34:6661–6672. doi: https://doi.org/10.1007/s00521-021-06058-8.
Canada Pork International (CPI). 1995. Canadian pork buyers’ manual. Canadian Pork International. Ontario, Canada.
Cuthbertson, H., G. Tar, K. Loudon, S. Lomax, P. White, P. McGreevy, R. Polkinghorne, and L. A. González. 2020. Using infrared thermography on farm of origin to predict meat quality and physiological response in cattle (Bos taurus) exposed to transport and marketing. Meat Sci. 169:108173. doi: https://doi.org/10.1016/j.meatsci.2020.108173.
de Souza, M. R., L. C. Ramos, R. de Oliveria Silva, E. M. dos Santos, E. da Silva Santos, E. O. da Silva, J. A. de Oliveria Silva, D. R. da Rocha, D. L. Hopkins, M. L. Chizzotti, and R. T. de Souza Rodrigues. 2025. Thermographic images for predicting the weight of lamb cuts. Anim. Prod. Sci. 65:AN25213. doi: https://doi.org/10.1071/AN25213.
Eriksson, D., and M. Poloczek. 2021. Scalable constrained Bayesian optimization. Proceedings of the 24th International Conference on Artificial Intelligence and Statistics, PMLR, 13–15 April 2021. Vol. 30: pp. 730–738.
Fernandes, A. F. A., J. R. R. Dórea, and G. J. de Magalhães Rosa. 2020. Image analysis and computer vision applications in animal sciences: an overview. Frontiers in Veterinary Science. 7:551269. doi: https://doi.org/10.3389/fvets.2020.551269.
Galanopoulos, K., S. Aggelopoulos, I. Kamenidou, and K. Mattas. 2006. Assessing the effects of managerial and production practices on the efficiency of commercial pig farming. Agr. Syst. 88:125–141. doi: https://doi.org/10.1016/j.agsy.2005.03.002.
Gygi, J. P., S. H. Kleinstein, and L. Guan. 2023. Predictive overfitting in immunological applications: pitfalls and solutions. Hum. Vacc. Immunother. 19:2251830. doi: https://doi.org/10.1080/21645515.2023.2251830.
Hasan, M. M., M. M. A. Chaudhry, C. Erkinbaev, J. Paliwal, S. P. Suman, and A. Rodas-Gonzalez. 2022. Application of Vis-NIR and SWIR spectroscopy for the segregation of bison muscles based on their color stability. Meat Sci. 188:108774. doi: https://doi.org/10.1016/j.meatsci.2022.108774.
Hawinkel, S., Waegeman, W., and S. Maere. 2024. Out-of-sample R2: estimation and inference. Am. Stat. 78:15–25. doi: https://doi.org/10.1080/00031305.2023.2216252.
Hawkins, D. M. 2004. The problem of overfitting. J. Chem. Inf. Comp. Sci. 44:1–12. doi: https://doi.org/10.1021/ci0342472.
Hein, T. 2020. Extreme risk for Canada’s hog producers due to COVID-19. Pig Progress. 5 August 2020. https://pigprogress.net/health-nutrition/extreme-risk-for-canadas-hog-producers-due-to-covid-19/. (Accessed 9 December 2024).https://pigprogress.net/health-nutrition/extreme-risk-for-canadas-hog-producers-due-to-covid-19/
Hite, L. M., A. D. Blair, K. R. Underwood, and J. K. Grubbs. 2020. Investigation on the use of a FLIR (forward looking infrared) camera to detect relationships between internal and surface temperatures of beef carcasses during chilling. South Dakota State University Department of Animal Science. https://openprairie.sdstate.edu/sd_beefday_2020/33. (Accessed 20 December 2025).https://openprairie.sdstate.edu/sd_beefday_2020/33
Hodnett, M., and J. F. Wiley. 2018. R Deep learning essentials: a step-by-step guide to building deep learning models using TensorFlow, Keras, and MXNet. Packt Publishing Ltd., Birmingham, UK.
Hurnik, J. F., A. B. Webster, and S. DeBoer. 1985. An investigation of skin temperature differentials in relation to estrus in dairy cattle using a thermal infrared scanning technique. J. Anim. Sci. 61:1095–1102. doi: https://doi.org/10.2527/jas1985.6151095x.
Kipper, M., M. Marcoux, I. Andretta, and C. Pomar. 2019. Assessing the accuracy of measurements obtained by dual-energy X-ray absorptiometry on pig carcasses and primal cuts. Meat Sci. 148:79–87. doi: https://doi.org/10.1016/j.meatsci.2018.10.005.
Leo, H., K. Saddami, Roslidar R. Muharar, K. Munadi, and F. Arnia. 2024. Lightweight convolutional neural network (CNN) model for obesity early detection using thermal images. Digital Health. 10:20552076241271639. doi: https://doi.org/10.1177/20552076241271639.
Lohumi, S., C. Wakholi, J. H. Baek, B. D. Kim, S. J. Kang, H. S. Kim, Y. K. Yun, W. Y. Lee, S. H. Yoon, and B.-K. Cho. 2018. Nondestructive estimation of lean meat yield of South Korean pig carcasses using machine vision technique. Korean J. Food Sci. An. 38:1109–1119. doi: https://doi.org/10.5851/kosfa.2018.e44.
Mali, S. 2026. Pork meat market analysis 2026. 8th Edition Cognitive Market Research. pp. 250. https://cognitivemarketresearch.com/pork-meat-market-report. (Accessed 4 January 2026).https://cognitivemarketresearch.com/pork-meat-market-report
Martins, B. M., A. L. C. Mendes, L. F. Silva, T. R. Moreira, J. H. C. Costa, P. P. Rotta, M. L. Chizzotti, and M. I. Marcondes. 2020. Estimating body weight, body condition score, and type traits in dairy cows using three dimensional cameras and manual body measurements. Livest. Sci. 236:104054. doi: https://doi.org/10.1016/j.livsci.2020.104054.
Masoumi, M., M. Marcoux, L. Maignel, and C. Pomar. 2021. Weight prediction of pork cuts and tissue composition using spectral graph wavelet. J. Food Eng. 299:110501. doi: https://doi.org/10.1016/j.jfoodeng.2021.110501.
Meisinger, D. J. 2003. The National Pork Quality benchmarking study. 56th Recip. Meat Conf. Proc., Columbia, MO 15–18 June 2003. pp. 71–72.
Miller, G. A., J. J. Hyslop, D. Barclay, A. Edwards, W. Thomson, and C.-A. Duthie. 2019. Using 3D imaging and machine learning to predict live weight and carcass characteristics of live finishing beef cattle. Frontiers in Sustainable Food Systems. 3:1–9. doi: https://doi.org/10.3389/fsufs.2019.00030.
Ngo, L., H. Ho, P. Hunter, K. Quinn, A. Thomson, and G. Pearson. 2016. Post-mortem prediction of primal and selected retail cut weights of New Zealand lamb from carcass and animal characteristics. Meat Sci. 112:39–45. doi: https://doi.org/10.1016/j.meatsci.2015.10.012.
Nguyen, V. 2019. Bayesian optimization for accelerating hyper-parameter tuning. 2019 IEEE Second International Conference on Artificial Intelligence and Knowledge Engineering (AIKE 2019). Sardinia, Italy 3–5 June 2019. pp. 302–305. doi: https://doi.org/10.1109/AIKE.2019.00056.
Nisbet, H., N. Lambe, G. Miller, A. Doeschl-Wilson, D. Barclay, A. Wheaton, C.-A. Duthie. 2024. Using in-abattoir 3-dimensional measurements from images of beef carcasses for the prediction of EUROP classification grade and carcass weight. Meat Sci. 209:109391. doi: https://doi.org/10.1016/j.meatsci.2023.109391.
National Research Council (NRC). 1998. Nutrient requirements of swine. 10th revised ed. The National Academies Press, Washington, DC. doi: https://doi.org/10.17226/6016.
OECD/FAO. 2025. OEC-FAO agricultural outlook 2025-2034. Organisation for Economic Co-operation and Development (OECD) and the Food and Agriculture Organization (FAO) of the United Nations. Paris and Rome. p. 164. doi: https://doi.org/10.1787/601276cd-en. (Accessed 6 January 2026).
Olfert, E. D., B. M. Cross, A. A. McWilliam, editors. 2009. Canadian Council on Animal Care Guide to the Care and Use of Experimental Animals. Vol. 1. 2nd ed. Ottawa, Canada.
Pei, X., Y. h. Zhao, L. Chen, Q. Guo, Z. Duan, Y. Pan, and H. Hou. 2023. Robustness of machine learning color, size change, normalization, and image enhancement on micrograph datasets with large sample differences. Mater. Design. 232:112086. doi: https://doi.org/10.1016/j.matdes.2023.112086.
Pomar, C., A. Fortin, and M. Marcoux. 2001. Estimation of meat yield and lean meat content (LMC) of pork carcasses using different methodologies for measuring fat thickness and back muscle. Journées de al Recherche Porcine. 33:71–77.
Pomar, C., and M. Marcoux. 2003. Comparing the Canadian pork lean yields and grading indexes predicted from grading methods based on Destron and Hennessy probe measurements. Can. J. Anim. Sci. 83:451–458. doi: https://doi.org/10.4141/A02-107.
R Core Team. 2023. R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/.https://www.R-project.org/
Ravi, N., V. Gabeur, Y.-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V. Alwala, N. Carion, C.-Y. Wu, R. Girshick, P. Dollár, and C. Feichtenhofer. 2024. SAM 2: segment anything in images and videos. arXiv. doi: https://doi.org/10.48550/arXiv.2408.00714.
Rius-Vilarrasa, E., L. Bünger, C. Maltin, K. R. Matthews, and R. Roehe. 2009. Evaluation of video image analysis (VIA) technology to predict meat yield of sheep carcasses on-line under UK abattoir conditions. Meat Sci. 82:94–100 doi: https://doi.org/10.1016/j.meatsci.2008.12.009.
Salamunesa, A. C. C., A. M. W. Stadnika, and E. B. Neves. 2017. The effect of body fat percentage and body fat distribution on skin surface temperature with infrared thermography. J. Therm. Biol. 66:1–9. doi: https://doi.org/10.1016/j.jtherbio.2017.03.006.
Schaefer, A. L., H. J. Perez Marquez, and C. J. Bench. 2025. The use of infrared thermography in the assessment of stress, health and metabolism in domestic animals. Can. J. Anim. Sci. 105:1–13. doi: https://doi.org/10.1139/cjas-2024-0067.
Schaefer, A. L., N. Cook, S. V. Tessaro, D. Deregt, G. Desroches, P. L. Dubeski, A. K. W. Tong, and D. L. Godson. 2004. Early detection and prediction of infection using infrared thermography. Can. J. Anim. Sci. 84:73–80. 23. doi: https://doi.org/10.4141/a02-104.
Schaefer, A. L., O. Iheshiulor, H. von Gaza, P. Charagu, G. Simpson, and A. Huisman. 2023. Thermal profiles: novel phenotypic measurements of animal growth and metabolic efficiency. J. Therm. Biol. 113:103537. doi: https://doi.org/10.1016/j.jtherbio.2023.103537.
Schaefer, A. L., N. J. Cook, J. S. Church, B. Basarb, B. Perry, C. Miller, and A. K. W. Tong. 2007. The use of infrared thermography as an early indicator of bovine respiratory disease complex in calves. Res. Vet. Sci. 83:376–384. doi: https://doi.org/10.1016/j.rvsc.2007.01.008.
Snekhalatha, U., K. P. Thanaraj, and K. Sangamithirai. 2021. Computer aided diagnosis of obesity based on thermal imaging using various convolutional neural networks. Biomed. Signal Proces. 63:102233. doi: https://doi.org/10.1016/j.bspc.2020.102233.
Soladoye, O. P., Ó. López Campos, J. L. Aalhus, C. Gariépy, P. Shand, and M. Juárez. 2016. Accuracy of dual-energy X-ray absorptiometry (DXA) in assessing carcass composition from different pig populations. Meat Sci. 121:310–316. doi: https://doi.org/10.1016/j.meatsci.2016.06.031.
Strecht, P., L. Cruz, C. Soares, J. Mendes-Moreira, and R. Abreu. 2015. A comparative study of classification and regression algorithms for modelling students’ academic performance. Proceedings of the 8th International Conference on Educational Data Mining, Madrid, Spain 26–29 June 2015.
Sun, D.-W. 2016. Computer vision technology for food quality evaluation. 2nd ed. Academic Press, Amersterdam. p. 635.
Taylor, W. E., D. C. Humphrey, B. D. Peyer, C. J. Cassady, S. M. Lonergan, and K. J. Stalder. 2023. Thermal imaging efficacy in predicting trim loss from cull sow carcasses. Meat and Muscle Biology. 7:16758, 1–10. doi: https://doi.org/10.22175/mmb.16758.
Usborne, W. R., D. Menton, and I. McMillan. 1987. Evaluation of the Destron PG-100 electronic probe for grading warm pork carcasses. Can. J. Anim. Sci. 67:209–212. doi: https://doi.org/10.4141/cjas87-024.
Zheng, S., M.-M. Cheng, J. Warrell, P. Sturgess, V. Vineet, C. Rother, and P. H. S. Torr. 2014. Dense semantic image segmentation with objects and attributes. 2014 Proc. CVPR IEEE. Columbus, OH 24–27 June 2014. pp. 3214–3221. doi: https://doi.org/10.1109/CVPR.2014.411.
Zhou, Z. Y., and B. M. Bohrer. 2019. Defining pig sort loss with a simulation of various marketing options of pigs with the assumption that marketing cuts improve variation in carcass weight and leanness. Can. J. Anim. Sci. 99: 542–552. doi: https://doi.org/10.1139/cjas-2018-0195.