Online evaluation of a commercial video image analysis system (Computer Vision System) to predict beef carcass red meat yield and for augmenting the assignment of USDA yield grades. United States Department of Agriculture

Visual Image Analysis for a new classification method of bovine carcasses according to EU legislation criteria

In the European Community, conformation and fat cover of bovine carcasses is assessed using the SEUROP grading system. In this study we pursued the development of an application software (App) based on Visual Image Analysis, useful for SEUROP and Fat Cover grading of bovine carcasses using a smartphone. The App was trained using 500 bovine carcasses. Carcass conformation and Fat Cover classes were assessed in parallel by expert evaluators and by App. Overall, a high correspondence was found between the measurements of carcasses parameters by operators and by the App, as high as 84.2% for SEUROP and 86.4% for the Fat Cover. In the 15.8% of samples with discordant SEUROP evaluation, and in the 13.6% of samples with discordant Fat Cover evaluation, the operators' and App measurements deviated by only one class. All values also aligned with the requirements expected by the current legislation for the use of automated and/or semi-automated systems able to determine the market value of carcasses.

Carcass and Primal Composition Predictions Using Camera Vision Systems (CVS) and Dual-Energy X-ray Absorptiometry (DXA) Technologies on Mature Cows

This study determined the potential of computer vision systems, namely the whole-side carcass camera (HCC) compared to the rib-eye camera (CCC) and dual energy X-ray absorptiometry (DXA) technology to predict primal and carcass composition of cull cows. The predictability (R²) of the HCC was similar to the CCC for total fat, but higher for lean (24.0%) and bone (61.6%). Subcutaneous fat (SQ), body cavity fat, and retail cut yield (RCY) estimations showed a difference of 6.2% between both CVS. The total lean meat yield (LMY) estimate was 22.4% better for CCC than for HCC. The combination of HCC and CCC resulted in a similar prediction of total fat, SQ, and intermuscular fat, and improved predictions of total lean and bone compared to HCC/CCC. Furthermore, a 25.3% improvement was observed for LMY and RCY estimations. DXA predictions showed improvements in R² values of 26.0% and 25.6% compared to the HCC alone or the HCC + CCC combined, respectively. These results suggest the feasibility of using HCC for predicting primal and carcass composition. This is an important finding for slaughter systems, such as those used for mature cattle in North America that do not routinely knife rib carcasses, which prevents the use of CCC.

Relationship between body size traits and carcass traits with primal cuts yields in Hanwoo steers

Objective

This study was conducted to evaluate the correlation between body size traits, carcass traits, and primal cuts in Hanwoo steers.

Methods

Sixty-one beef carcasses were classified for conformation and primal cut weight. Additionally, carcass weight, fat thickness, carcass dimensions, and longissimus muscle area were determined to complement the grading.

Results

The average live weight and cold carcass weight were 759 and 469 kg, respectively. The mean carcass meat, fat, and bone proportions were 551, 298, and 151 g/kg, respectively. Primal cuts weights showed significant positive correlations (p<0.001) of 0.42 to 0.82 with live weight, carcass weight, and longissimus muscle area and a significant negative correlation with carcass fat (without shank, −0.38 to −0.10). Primal cut weights were positively correlated (p<0.01) with carcass length (0.41 to 0.77), forequarter length (0.33 to 0.57), 6th lumbar vertebrae–heel length (0.33 to 0.59), 7th cervical vertebrae carcass breadth (0.35 to 0.58), 5th to 6th thoracic vertebrae breadth (0.36 to 0.65), 7th to 8th thoracic vertebrae girth (0.38 to 0.63), and coxae girth (0.34 to 0.56) and non-significantly related to cervical vertebrae length and coxae thickness.

Conclusion

There was a high correlation among live weight, carcass weight, longissimus muscle area, carcass length, 7th cervical vertebrae carcass breadth, 5th to 6th thoracic vertebrae breadth, and 7th to 8th thoracic vertebrae girth of the primal cuts yield. The correlation between fat and primal cut yields was highly significant and negative. Carcass length and 7th to 8th thoracic vertebrae girth, appear to be the most important traits affecting primal cut yields.

Opportunities to improve the accuracy of the United States Department of Agriculture beef yield grade equation through precision agriculture1

The U.S. Department of Agriculture (USDA) yield grade (YG) equation is used to predict the retail yield of beef carcasses, which facilitates a more accurate payment for cattle when they are sold on a grid pricing system that considers carcass composition instead of body weight alone. The current USDA YG equation was developed over 50 yr ago. Arguably, the population of cattle used to develop the YG equation is different than the current diverse U.S. beef cattle supply today. The objectives of this manuscript are to promote the adoption and use of precision agriculture technologies (i.e., camera grading and electronic animal identification) throughout the U.S. beef supply chain as a means to enhance the ability of the USDA YG equation to more accurately predict the retail yield across the population of cattle that contributes to the current U.S. beef supply. Camera grading has improved the accuracy of determining beef carcass retail yield; however, the use of electronic animal identification would allow for additional information to be passed back and forth between the packer, cattle feeder, and producer. Information, such as sex, genetics, medical treatment history, diets consumed, and growth promotant administration, as well as other information could be used to create additional variables for a new augmented USDA YG equation. Herein, fabrication yields demonstrated a 5.6 USDA YG and 12.8% boneless closely trimmed retail cut difference between actual cutout measurements and calculated values from the USDA YG equation for Jersey-influenced cattle. Evidence of such disparities between calculated and actual values warrants a reevaluation of the USDA YG system and consideration for implementing advancements in precision agriculture to improve the prediction of beef carcass retail yield to more accurately account for the large amount of variation in beef carcass retail yield from the cattle in the United States.

A Potential Use of 3-D Scanning to Evaluate the Chemical Composition of Pork Meat

The aim of this study was to determine the possibility of 3-D scanning method in chemical composition evaluation of pork meat. The sampling material comprised neck muscles (1000 g each) obtained from 20 pork carcasses. The volumetric estimation process of the elements was conducted on the basis of point cloud collected using 3-D scanner. Knowing the weight of neck muscles, their density was calculated which was subsequently correlated with the content of basic chemical components of the pork meat (water, protein and fat content, determined by standard methods). The significant correlations (P ≤ 0.05) between meat density and water (r = 0.5213), protein (r = 0.5887), and fat (r = -0.6601) content were obtained. Based on the obtained results it seems likely to employ the 3-D scanning method to compute the meat chemical composition.

PRACTICAL APPLICATION

The use of the 3-D scanning method in industrial practice will allow to evaluate the chemical composition of meat in online mode on a dressing and fabrication line and in a rapid, noninvasive manner. The control of the raw material using the 3-D scanning will allow to make visual assessment more objective and will enable optimal standardization of meat batches prior to processing stage. It will ensure not only the repeatability of product quality characteristics, but also optimal use of raw material-lean and fat meat. The knowledge of chemical composition of meat is essential due to legal requirements associated with mandatory nutrition facts labels on food products.

Accuracy of body measurements using digital image analysis in female Holstein calves

The objective of this study was to determine the accuracy of body measurements (BM) in Holstein female calves using digital image analysis. BM including body length, wither height, chest depth, hip height, and hip width of calves were recorded by stick and tape measurements at birth, weaning and 24 weeks of age. Then photos of calves were taken while calves were standing in a squeeze chute by a digital camera and were analysed by image analysis software to obtain BM of each calf from the image in centimetres. After comparing the actual and predicted BM, the accuracy was determined as 71, 97 and 99% for body length, 69, 87 and 99% for wither height, 43, 98 and 99% for chest depth, 74, 99 and 99% for hip height and 53, 99 and 98% for hip width at birth, weaning and 24 weeks of age, respectively. The difference between actual and predicted BM was significant at birth (P < 0.01). Although there were numerical differences between actual and predicted BM, the differences were not significant at weaning and 24 weeks of age (P > 0.01). According to these results, the BM estimation of Holstein female calves using digital image analysis produced high prediction accuracy at weaning and 24 weeks of age, but not at birth. The data presented in this study indicate that the digital image analysis provides very close agreement and reality for prediction of BM of Holstein female calves.

An industry applicable model for predicting lean meat yield in lamb carcasses

A wide selection of lamb types (n = 360) of mixed sex (ewes and wethers) were slaughtered at a commercial abattoir. Soft tissue depth at the GR site (thickness of tissue over the 12th rib, 110 mm from the midline) was measured in the chiller, using a GR knife (GR). Each carcass was subsequently broken down to a range of trimmed boneless retail cuts and the lean meat yield determined. The predominant industry model for predicting meat yield in Australia uses hot carcass weight (HCW) and tissue depth at the GR site. A moderate level of accuracy and precision was found when HCW and GR were used to predict lean meat yield (R2 = 40.5, r.s.d. = 2.39%), which could be improved markedly when loin muscle cross-sectional area at the 12th rib (EMA) was included in the model (R2 = 54.5, r.s.d. = 2.10%). A better result was achieved when the model included the weight of subcutaneous fat (SLFat) from the shortloin (R2 = 73.8, r.s.d. = 1.59%). A combination of SLFat and the weight of the shortloin muscle (SLMus) negated the need to include either GR or EMA in the model (R2 = 76.1, r.s.d. = 1.52%). The transportability of a model based on HCW, SLFat and SLMus was tested by randomly dividing the dataset and comparing the coefficients and the level of accuracy and precision. Collecting measures of EMA, SLFat and SLMus in boning rooms is potentially feasible. If this can be achieved under commercial conditions, a rigorous method for automatically predicting lean meat yield during boning could be applied. Application of the approach to large-scale research programs, where estimates of lean meat yield are required, would be possible at a reduced cost compared with alternative systems based on full carcass breakdown. A suitable model is given for this purpose.

Genetic correlations between live yearling bull and steer carcass traits adjusted to different slaughter end points. 1. Carcass lean percentage

We studied genetic relationships between age-constant live yearling beef bull growth and ultrasound traits and steer carcass traits with dissected steer carcass lean percentage adjusted to slaughter age-, HCW-, fat depth-, and marbling score-constant end points. Three measures of steer carcass lean percentage were used. Blue Tag lean percentage (BTLean) was predicted from HCW, fat depth, and LM area measurements. Ruler lean percentage (RulerLean) was predicted from carcass fat depth and LM depth and width measurements. Dissected lean percentage (DissLean) was based on dissection of the 10-11-12th rib section. Both BTLean (h2 = 0.30 to 0.44) and DissLean (h2 = 0.34 to 0.39) were more heritable than RulerLean (h2 = 0.05 to 0.14) at all end points. Genetic correlations among DissLean and RulerLean (rg = 0.61 to 0.70), DissLean and BTLean (rg = 0.56 to 0.72), and BTLean and RulerLean (rg = 0.59 to 0.90) indicated that these traits were not genetically identical. Adjusting Diss-Lean to different end points changed the magnitude, but generally not the direction, of genetic correlations with indicator traits. Ultrasound scan-age-constant live yearling bull lean percentage estimates were heritable (h2 = 0.26 to 0.42) and genetically correlated with each other (rg = 0.68 to 0.99) but had greater correlations with DissLean at slaughter age (rg = 0.24 to 0.48) and HCW (rg = 0.16 to 0.40) end points than at fat depth (rg = -0.08 to 0.13) and marbling score (rg = 0.02 to 0.11) end points. Scan-age-constant yearling bull ultrasound fat depth also had stronger correlations with DissLean at slaughter age (rg = -0.34) and HCW (rg = -0.25) than at fat depth (rg = -0.02) and marbling score (rg = -0.03) end points. Yearling bull scan-age-constant ultrasound LM area was positively correlated with DissLean at all endpoints (rg = 0.11 to 0.23). Genetic correlations between yearling bull LM method 1 width (rg = 0.38 to 0.56) and method 2 depth (rg = -0.17 to -0.38) measurements with DissLean suggested that LM shape may be a valuable addition to genetic improvement programs for carcass lean percentage at slaughter age, HCW, and fat depth constant end points. At all end points, steer carcass fat depth (rg = -0.60 to -0.64) and LM area (rg = 0.48 to 0.59) had stronger associations with DissLean than did corresponding live yearling bull measurements. Improved methods that combine live ultrasound and carcass traits would be beneficial for evaluating carcass lean percentage at fat depth or marbling score end points.

An evaluation of the lamb vision system as a predictor of lamb carcass red meat yield percentage

An objective method for predicting red meat yield in lamb carcasses is needed to accurately assess true carcass value. This study was performed to evaluate the ability of the lamb vision system (LVS; Research Management Systems USA, Fort Collins, CO) to predict fabrication yields of lamb carcasses. Lamb carcasses (n = 246) were evaluated using LVS and hot carcass weight (HCW), as well as by USDA expert and on-line graders, before fabrication of carcass sides to either bone-in or boneless cuts. On-line whole number, expert whole-number, and expert nearest-tenth USDA yield grades and LVS + HCW estimates accounted for 53, 52, 58, and 60%, respectively, of the observed variability in boneless, saleable meat yields, and accounted for 56, 57, 62, and 62%, respectively, of the variation in bone-in, saleable meat yields. The LVS + HCW system predicted 77, 65, 70, and 87% of the variation in weights of boneless shoulders, racks, loins, and legs, respectively, and 85, 72, 75, and 86% of the variation in weights of bone-in shoulders, racks, loins, and legs, respectively. Addition of longissimus muscle area (REA), adjusted fat thickness (AFT), or both REA and AFT to LVS + HCW models resulted in improved prediction of boneless saleable meat yields by 5, 3, and 5 percentage points, respectively. Bone-in, saleable meat yield estimations were improved in predictive accuracy by 7.7, 6.6, and 10.1 percentage points, and in precision, when REA alone, AFT alone, or both REA and AFT, respectively, were added to the LVS + HCW output models. Use of LVS + HCW to predict boneless red meat yields of lamb carcasses was more accurate than use of current on-line whole-number, expert whole-number, or expert nearest-tenth USDA yield grades. Thus, LVS + HCW output, when used alone or in combination with AFT and/or REA, improved on-line estimation of boneless cut yields from lamb carcasses. The ability of LVS + HCW to predict yields of wholesale cuts suggests that LVS could be used as an objective means for pricing carcasses in a value-based marketing system.

Evaluation of the E+V video image analysis system as a predictor of pork carcass meat yield

This study was conducted to assess the ability of the VCS2001 (E+V, Oranienburg, Germany) video image analysis system to predict pork carcass composition. Pork carcasses (n = 278) were selected from a commercial packing plant to differ in weight, Fat-O-Meater (FOM) predicted percentage lean, and gender. Carcasses were imaged three times with the VCS2001, chilled overnight, and then sequentially fabricated into boneless subprimals. The VCS2001 accurately predicted the weight of total saleable product (R2 = 0.88, root mean square error [RMSE] = 1.84) and fat-corrected lean (R2 = 0.92, RMSE = 1.66), but autocorrelation existed between dependent and independent variables. The VCS2001 was acceptably accurate and precise in predicting weights of bone-in ham (R2 = 0.83, RMSE = 0.80), bone-in loin (R2 = 0.74, RMSE = 1.17), loin lean (R2 = 0.77, RMSE = 0.82), belly (R2 = 0.78, RMSE = 0.94), sparerib (R2 = 0.55, RMSE = 0.28), and boneless shoulder (R2 = 0.73, RMSE = 0.79). Weights were more accurately predicted than yields (as a percentage of hot carcass weight) of total saleable product (R2 = 0.47, RMSE = 1.97) or total fat-corrected lean (R2 = 0.44, RMSE = 1.89) using VCS2002, and it did not accurately predict percentages of bone-in ham (R2 = 0.45, RMSE = 1.13), ham lean (R2 = 0.32, RMSE = 1.46), bone-in loin (R2 = 0.29, RMSE = 1.36), loin lean (R2 = 0.56, RMSE = 0.90), belly (R2 = 0.43, RMSE = 1.08), sparerib (R2 = 0.08, RMSE = 0.32), or boneless shoulder (R2 = 0.30, RMSE = 0.88). New prediction models and equations were developed using VCS2001 output variables plus hot carcass weight to predict weight of total saleable product (R2 = 0.89, RMSE = 1.72) and fat-corrected lean (R2 = 0.93, RMSE = 1.55) with very minimal increases in accuracy and precision over that achieved using E+V-programmed models and equations. Use of new prediction models and equations marginally improved accuracy and precision of estimations of total saleable product yield (R2 = 0.56, RMSE = 1.81) and fat-corrected lean yield (R2 = 0.57, RMSE = 1.67) over that achieved using E+V-programmed models and equations. The VCS2001 was not able to predict pork carcass composition more accurately than existing technology; therefore, further development is needed to assure commercial viability of this instrument.

Online prediction of beef tenderness using a computer vision system equipped with a BeefCam module

Four experiments were conducted in two commercial packing plants to evaluate the effectiveness of a commercial online video image analysis (VIA) system (the Computer Vision System equipped with a BeefCam module [CVS BeefCam]) to predict tenderness of beef steaks using online measurements obtained at chain speeds. Longissimus muscle (LM) samples from the rib (Exp. 1, 2, and 4) or strip loin (Exp. 3) were obtained from each carcass and Warner-Bratzler shear force (WBSF) was measured after 14 d of aging. The CVS BeefCam output variable for LM area, adjusted for carcass weight (cm2/kg), was correlated (P < 0.05) with WBSF values in all experiments. The CVS BeefCam lean color measurements, a* and b*, were effective (P < 0.05) in all experiments for segregating carcasses into groups that produced LM steaks differing in WBSF values. Fat color measurements by CVS BeefCam were usually ineffective for segregating carcasses into groups differing in WBSF values; however, in Exp. 4, fat b* identified a group of carcasses that produced tough LM steaks. Quality grade factors accounted for 3, 18, 21, and 0% of the variation in WBSF among steaks in Exp. 1 (n = 399), 2 (n = 195), 3 (n = 304), and 4 (n = 184), respectively, whereas CVS BeefCam output variables accounted for 17, 30, 19, and 6% of the variation in WBSF among steaks in Exp. 1, 2, 3, and 4, respectively. A multiple linear regression equation developed with data from Exp. 2 accurately classified carcasses in Exp. 1 and 4 and thereby may be useful for decreasing the likelihood that a consumer would encounter a tough (WBSF > 4.5 kg) LM steak in a group classified as "tender" by CVS BeefCam compared with an unsorted population. Online measurements of beef carcasses by use of CVS BeefCam were useful for predicting the tenderness of beef LM steaks, and sorting carcasses using these measurements could aid in producing groups of beef carcasses with more uniform LM steak tenderness.

Effectiveness of the SmartMV prototype BeefCam System to sort beef carcasses into expected palatability groups

This study was conducted to determine the effectiveness of the SmartMV prototype BeefCam Video Imaging System (prototype BeefCam) for classifying beef carcasses into palatability ("certified" or "not certified" as palatable) groups. Carcasses (n = 769) were selected from four beef-packing plants to represent three USDA quality grade groups (Top Choice, TC; Low Choice, LC; and Select, SE). Following chilling, a prototype BeefCam image of the longissimus muscle was obtained for each carcass. Strip loins were collected from the left side of each carcass and aged for 10 d; Warner-Bratzler shear force (WBSF; n = 769) values and consumer panel ratings (hedonic, end-anchored, 9-point ratings for overall like/dislike; n = 500 carcasses) were obtained for cooked steaks. Using information from the images, two regression models were developed to predict the first principal component of WBSF and consumer panel ratings for sorting carcasses based on expected eating quality. Model I used only prototype BeefCam output, whereas Model II used prototype BeefCam output and a coded value for quality grade group. For both models, carcasses with a predicted value of less than 0.0 were certified as producing palatable beef Additional carcasses (n = 292) were evaluated at a fifth and separate packing plant by prototype BeefCam to validate Models I and II. A strip loin was collected from each carcass and WBSF was measured after 14 d of aging. The percentages of validation carcasses that generated tough (WBSF > or = 4.5 kg) steaks were 6.5,5.8,10.7, and 7.9% for, TC, LC, SE, and all carcasses, respectively. Use of Model I certified 51.9, 47.6, 43.8, and 47.3% of TC, LC, SE, and all carcasses, respectively. Of the carcasses certified by use of Model I, 0.0,0.0, 4.1, and 1.4% of TC, LC, SE, and all carcasses, respectively, generated tough steaks. Use of Model II certified 59.7, 47.6, 25.0, and 42.1% of TC, LC, SE, and all carcasses, respectively. Of the carcasses certified by use of Model II, 2.2, 0.0, 3.6, and 1.6% of TC, LC, SE, and all carcasses, respectively, generated tough steaks. For both models, the frequency of carcasses that produced tough steaks in the certified group was lower (P < 0.05) for all validation carcasses sampled compared with that of the original carcass population. Based on the decrease in the frequency of carcasses that produced tough steaks, further development of a commercial BeefCam system is warranted.