Utilization of image processing to quantitate surface metmyoglobin on fresh beef

Meat Scknw, Vol. 43,Nos 3 4, 265 214.1996

PII:

s0309-1740(95)0006l-5

Copyright ‘9’ 1996 Elsevier Science Ltd Printed in Great Britain. All rights reserved 0309.1740!96 $15.00+ 0.00

ELSEVIER

Utilization of Image Processing to Quantitate Surface Metmyoglobin on Fresh Beef B. P. Demoq”

D. E. Gerrard,h

X. Gao,” J. Tan” & R. W. Mandigod*

uArmour Swift-Eckrich, Downers Grove, IL, USA of Food Science and Human Nutrition, Department of Animal Science, University of Missouri-Columbia, Columbia, MO 65211, USA “Department of Agricultural Engineering, University of Missouri-Columbia, Columbia, MO 65211, USA “Department of Animal Science, University of Nebraska-Lincoln, Lincoln, NE 68583-0908, “Department

(Received

I2 May 1995; accepted

I 1 September

USA

1995)

ABSTRACT Ground betf patties were mangfactured with various combinations of ascorbic acid and mechanically recovered neck bone lean (MRNL) to study the use of image processing in predicting percentage surface metmyoglobin (metMb) on fresh beef. Ascorbic acid and MRNL cause various color phenomena that resulted in a wide range of variation in surface color. Patties were also stored over si.x days of retail display to cause further color changes. Surface color was assessed by several different accepted methods. A prediction equation for percentage surface metMb included mean valuesfor hue, saturation and intensity. Root mean square error, R-square and Mallow’s Cp statistic were used as selection criteria for choosing the best predictive model. Image processing hue, saturation and intensity accountedfor 93% of the variation in percentage surface metMb. Since hue, saturation and intensity each contribute to overall color projile it is logical that these parameters are good predictors. These data indicate that image processing is capable qf objectively measuring percentage surface metMb. Copyright 10 1996 Elsevier Science Ltd INTRODUCTION Discoloration due to pigment oxidation at the surfaces of meat is usually measured by assessing the relative concentration of metmyoglobin (metMb). Accumulation of metMb is responsible for the undesirable brown color that forms on meat surfaces during shortterm fresh retail display. There is a wealth of information available on quantitating surface metMb, however, there is still debate as to the reliability of the methods in use. Among the methods in use to quantitate surface metMb, the most controversial are those that involve pigment extraction. Extraction procedures come under sharp criticism because they are time consuming, destroy the sample analysed and likely result in oxidative and reductive changes in pigments (Stewart et al., 1965). Despite these criticisms, Chen & Trout (1991) and Demos & Mandigo (1996) reported acceptable results with these methods. It is important to control pH and temperature if these methods must be used. *To whom correspondence

should be addressed. 265

B. P. Demos et al.

266

Another common method used to quantitate surface metMb is based on the calculation of K/S ratios (Stewart et al., 1965). MetMb is determined by reflectance data from the ratio K/S 525 rim/K/// 572 nm. Limiting values for this ratio are established for meat containing 100% and 0% metMb, and a linear relation is assumed between the ratios and intermediate amounts of metMb. This method also has been criticized. Kryzwicki (1979) believes that meat samples cannot be completely converted to 100% or 0% metMb and that colorless structural elements in meat cause errors in K/S ratios that do not occur in uniformly pigmented materials. Arguably the most accurate method currently available is the method of Kryzwicki (1979). This method allows the amount of metMb to be calculated without the need to determine limiting values for 100% and 0% metMb on the surface and instead takes into account only the molar absorbance coefficients of metMb. Recent work with image processing has allowed researchers to determine fat percentage of ground beef as well as quality factors of intact muscle due to color differences of samples (Gerrard ef al., 1995; Gwartney et al., 1996). This technology has the advantage of capturing the image of an entire sample (e.g. a ground beef patty, a pork loin chop, etc.) as opposed to evaluating only a small area as is the case with virtually all other methods. The objective of this study was to determine the feasibility of quantitating surface metMb on ground beef patties with image processing. MATERIALS

AND METHODS

Product formulation

Lean and fat beef trim from USDA Select and Standard carcasses was obtained seven days post mortem from the University of Nebraska Loeffel Meat Laboratory. All trim was coarse ground (Hobart Manufacturing Co., Troy, OH) through a 2.54 cm plate, vacuum packaged and frozen at -35°C for 14 days. Mechanically recovered neck bone lean was manufactured as described by Demos & Mandigo (1995a) and frozen to -35°C. Grab samples were taken for fat determination by ether extraction (AOAC, 1990). All raw materials were tempered 24 hr at 4°C. Lean and fat beef trim, MRNL and dry ascorbic acid (Asc, Fisher Scientific, Fair Lawn, NJ) were combined in the appropriate ratios to yield the following combinations: 0 ppm Asc/O% MRNL, 0 ppm Asc/lS% MRNL, 0 ppm Asc/30% MRNL, 1000 ppm Asc/O% MRNL, 1000 ppm Asc/lS% MRNL, 1000 ppm Asc/30% MRNL. All combinations were formulated to 20% fat. Each 11.4 kg formulation was mixed for 5 min (Model 100 DA Food Mixer, Leland Detroit Manufacturing Co., Detroit, MI) and ground through a 4.7 mm plate. Quarter pound patties were processed with a Hollymatic patty machine (Model 580, Hollymatic Corp., Park Forest, IL). Patties were placed on Styrofoam trays, two to a tray, and overwrapped with polyvinyl chloride. Retail storage and sampling

Overwrapped patties were stored at 7°C under continuous warm white fluorescent illumination (700 lux) 5 days (days &6, day of manufacture = day 0). Patties were selected at random and measured for surface discoloration on each day of display according to the procedures that follow. Surface metmyoglobin

concentration-Kryzwicki

method

A modified method of Kryzwicki (1979) was used to quantitate percentage surface metmyoglobin. Surface reflectance of patties was measured with a HunterLab calorimeter

Utilization of image processing

267

(LabScan II, Hunter Associates Laboratory, Inc., Reston, VA), equipped with a 30 mm viewing port. The port was covered with polyvinyl chloride to prevent drippings from entering. The polyvinyl chloride did not alter readings to any measurable extent. Three reflectance readings of three patties per day/replication combination were read from 40& 700 nm. Reflectance readings were converted to absorbance [2 - log (% reflectance)] and used in the following equation: metmyoglobin( CIE L* a* b* values were obtained

%) = 1.395 -

(A572 - A700) x 1oo (A525 - A700)

from the same readings.

Image processing The image processing system used for this experiment consisted of a Sony XC-71 1 CCD camera, a Sony PVM-13424 color video monitor, a Data Translation DT2871 color image frame grabber, a DT2878 advanced processor, AURORA and AIPL programming libraries and 486-50 Hz microcomputer for the development of processing software. The system was programmed in Microsoft C/C+ + 7.00. Ground beef sample images were taken individually with a uniform non-glare black background. The color images were 512x483 pixels in resolution. Each pixel represented an actual area of 0.134 mm’. The images were represented in the HSI (hue, saturation and intensity) and RGB (red, green, blue) format. Background segmentation was first performed on the original images to give a uniformly black background. Features characterizing color of the patties were extracted from the image. The histograms of H, S and I were approximately normal. Statistics of these histograms were used as the color features, which were: Mean: Mode: Skew3: Sk3:

uH, US and u1 value of most frequent occurrence third moment of the mean third moment of the mode where the subscripts

Experimental

denote

the color components.

design and data analysis

Statistical analysis included linear regression using the Statistical Analysis System (SAS, 1991). To determine the best predictive model, all possible regression models for 12 independent variables were evaluated with correlation coefficients, R-squares, root mean square error (RMSE) and Mallow’s Cp statistic (Mallows, 1973). The prediction model selected had maximum R-square, minimum RMSE and a Cp statistic closest to the number of parameters in the model. The RMSE is an indicator of deviation of the predicted value from the regression line. Mallow’s Cp statistic is a measure of bias that may exist in the prediction model.

RESULTS There are two fresh beef. One blue (R, G and distribution of

AND

DISCUSSION

ways that image processing can be useful in measuring surface color of way is to get an indication of the three primary colors: red, green and B, respectively). When the image of the ground beef patty is captured, a computer pixels is obtained for each of these colors. Since fresh beef is

B. P. Demos et al.

268

R*=.96 RMSE=1.48

15

10 130

140

150

180

170

180

190

Red value

Fig. 1. CIE a* vs image processing red value for ground beef patty surfaces.

normally bright cherry red, it is logical that R values would be the most valuable indicator of fresh meat. Currently the most common method of measuring surface redness is the CIE a* value. The a* value is a measure of a color continuum from red to green; a higher number is indicative of a more red color. Therefore, it is not surprising that CIE a* values, as measured with a HunterLab calorimeter, were highly correlated to image processing R mean values (JJ < 0.01, Y= 0.98). Figure 1 shows a* values vs R-values. From these data it is apparent that image processing R values can be used to measure surface redness of ground beef. Theoretically, R mean values should be a better assesment of surface redness because the entire patty is analyzed compared to only a small area, as is done with the HunterLab calorimeter. Since a* values are on a continuum from red to green it is logical to expect a negative correlation between image processing G values and a* values (JJ< 0.01, Y= -0.96). Figure 2 shows that this negative relationship occurred in our study. The same relationship was observed for b’, a continuum from yellow to blue, and image processing B mean values (Fig. 3). A higher b* value is indicative of greater yellowness, a lower number is indicative of greater blueness. There is a strong negative correlation between CIE b* values and image processing B mean values ((p
0.99 -0.13 -0.13 -0.79 -0.76 0.25 -0.20 0.87 0.79 0.40 -0.25

0.94 -0.09 -0.09 -0.76 -0.74 0.23 -0.16 0.76 0.75 0.52 -0.19

lnt = intensity.

1.0

Huemean

1.0 -0.17 -0.17 -0.76 -0.73 0.21 -0.21 0.85 0.76 0.38 -0.24 1.0 1.0 -0.09 -0.12 0.08 0.13 -0.04 0.05 0.13 -0.09

Hueskew

1.0 -0.10 -0.13 0.09 0.12 -0.03 0.06 0.14 -0.09

Huesk

Surface Metmyoglobin

Huevmo

for Percentage

1.0 0.94

Met(%)

Coefficients

Sat = Saturation,

Met (%) Hue mean Hue vmo Hue skew Hue sk Sat mean Sat vmo Sat skew Sat sk Int mean Int vmo Int skew Int sk

Correlation

-0.24 0.31 -0.92 -0.90 -0.55 0.32

I.0 I.0

Satmean

1.0 -0.23 0.32 -0.89 -0.90 -0.55 0.30

Satvmo

TABLE 1 and Image Processing

1.0 0.55 0.20 0.20 0.09 -0.27

Satskew

Variables

1.0 0.88 0.44 -0.33

-0.33 -0.01 -0.27

Intmean

1.0 0.35 -0.20

Intvmo

in Developing

-0.30

I.0

Satsk

Considered

1.0 -0.31

Intskew

a Prediction

1.0

In tsk

Equation

2 d Z‘

c

B. P. Demos et al.

270

1 R2=.92 RMSE=2.20

10 40

I

I

I

I

I

I

50

60

70

60

90

100

Green

Fig. 2. CIE a* vs image processing

110

value

green value for ground beef patty surfaces.

surface metMb would ultimately make image processing the method of choice for evaluating fresh beef color because of its ability to measure the entire surface of a meat cut. In order to achieve this, the most accurate method of quantitating surface metMb that is currently available had to be determined. Based on earlier findings (Demos & Mandigo, 19956) and this experiment, it was determined that the method of Kryzwicki (1979) was the most accurate. The Kryzwicki method routinely has the smallest standard errors and has advantages over the other two methods as discussed in the introduction. In addition to values for R, G and B, image processing is capable of generating hue, saturation and intensity data. A distribution of computer pixels is obtained for each of these components. Four variables are then generated from each distribution: mean; mode; skew3 and sk3, resulting in a total of 12 variables. Mean is the average for each color component and skew3 and sk3 are measures of distribution skew. In each case, the distribution was normal, so mean is likely the best predictor from each distribution and all other variables are less informative. None-the-less, all variables were analyzed with PROC RSQUARE of SAS (1991) to evaluate the best predictors of surface metMb. Mallow’s Cp statistic, R-square and root mean square error were used as selection criteria to determine the best 3-variable model that would predict percentage surface metMb. The best 3-variable model included the mean hue, mean saturation and mean intensity values. Since each of these components adds something different to the overall color profile and all distributions were normal, it is logical that these three variables would turn out to be the best predictors. In each case, the mean value for each of the predictors had an equal or higher correlation to surface metMb than the mode, skew3 or sk3 values (Table 1). Percentage surface metMb vs hue mean, saturation mean and intensity mean are shown in Figs 4, 5 and 6, respectivley. Hue mean by itself appears to be the best sole predictor of surface metMb (Fig. 4). In contrast, saturation mean and intensity mean individually show a curvilinear relationship to percentage surface metMb and the R*s are weaker than for hue mean vs percentage surface metMb (Figs 5 and 6). When all three

Utilization of image processing

271

R2=.79 RMSE=i.W

10

20

30

40

50

60

70

Blue value

Fig. 3. CIE b’ vs image processing

blue value for ground

beef patty surfaces.

70 R’=.SS RMSE=4.11% 60

20 10

I 0

5

10

15

20

25

Hue mean

Fig. 4. Percentage

surface

metmyoglobin

vs image processing surfaces.

hue mean

for ground

beef patty

B. P. Demos et al.

272 70

R2=.57

“h

60

RMSE=7.96%

20

10 100

120

140

160

Saturation

Fig. 5. Percentage

surface metmyoglobin

160

200

mean

vs image processing surfaces.

saturation

mean for ground beef patty

70 R2=.56 RMSE=7.07%

60

.

.

. . .

.

n

n

n n

_

00

90

intensity

Fig. 6. Percentage

surface metmyoglobin

mean

vs image processing surfaces.

intensity

mean for ground

beef patty

273

Utilization qf image processing

R2=.93 RMSE=3.16%

10

’

/

I

10

20

30 Actual

Fig. 7.

Actual

vs

,

40 percentage

surface

I

I

I

50

60

70

metmyoglobin

predicted percentage surface metmyoglobin

variables are used together to predict square error improve significantly.

percentage

surface

for ground beef patty surfaces.

metMb,

R* and

root

mean

The final prediction equation is as follows: metMb(%)

= 213.27 + 3.89(hue)

- 0.24(sat)

- 2.07(int)

(R* = 0.93)

Other prediction equations that result in higher R-squares are possible if more variables are utilized, but improvement in R-square is minimal and the most logical, practical model includes the mean values for hue, saturation and intensity. The quadratic terms for hue mean, saturation mean and intensity mean were tested to see if the model could be improved, but improvement was negligible and the terms were dropped. As a final test of the prediction equation, actual percent metMb was plotted against predicted metMb (Fig. 7). Wide variation in samples was intended in this study, so that an accurate representation of surface discoloration would be obtained. This creates a solid test of the new method over a broad range of surface colors. It is apparent that image processing is capable of predicting percentage surface metMb over a broad range of surface colors.

CONCLUSIONS Image processing is a new technology that is an effective tool for use in research situations for assessing percentage surface metMb on fresh beef. The predictions for the varying surface colors were accurate, as demonstrated by the high R-square and RMSE. Image processing is effective on a heterogenous population of surface colors.

214

B. P. Demos et al REFERENCES

AOAC. (1990) OfJiciul Methods of Analysis, 15th Edn. Assoc. of Official Analytical Chemists, Washington, DC. Chen, C. M. & Trout, G. R. (1991). J. Food Sci., 56, 1461. Demos, B. P. & Mandigo, R. W. (1995~). Meat Sci., 60, 576. Demos, B. P. & Mandigo, R. W. (1996). Meat Sci., 42, 415. Gerrard, D. E., Gao, X. & Tan, J. (1995). J. Food Sci. (submitted). Gwartney, B. L., Gerrard, D. E., Gao, X. & Tan, J. (1996). J. Must. Foods (in press). Kryzwicki, K. (1979). Meat Sci., 3, 1. Mallows, C. L. (1973). Some comments on Cp. Technometrics, 15, 661. SAS (1991). SASjSTAT User’s Guide, Release 6.03, SAS Institute, Cary, NC. Stewart, M. R., Hutchins, B. K., Zipser, M. W. & Watts, B. M. (1965). J. Food Sci., 30, 487.