COMPUTER VISION SYSTEM FOR COLOR MEASUREMENTS OF MEAT AND MEAT PRODUCTS: A REVIEW

Meat and meat products color evaluation ability of a computer vision system (CVS) is investigated by a comparison study with color measurements from a traditional colorimeter. A statistical analysis revealed significant differences between the instrumental values in all three dimensions (L*, a*, b*) between the CVS and colorimeter. The CVS-generated colors were more similar to the sample of the meat products visualized on the monitor, compared to colorimeter-generated colors in all (100 %) individual trials performed. The use of CVS should be considered a superior alternative to the traditional method for measuring color of meat and meat products. УДК /UDC: 637.5.04/.07:004.93 DOI 10.21323/2414–438X-2018–3–4–4–15 ДЛЯ ЦИТИРОВАНИЯ: Томашевич И.Б. Cистема компьютерного зрения для измерения цветовых параметров мяса и мясных продуктов: Обзор. Теория и практика переработки мяса. 2018; 4(4): 4–15. DOI 10.21323/2414–438X‐2018–3–4–4–15 FOR CITATION: TomaševićI. B. Computer vision system for color measurements of meat and meat products: A review. Theory and practice of meat processing. 2018;4(4): 4–15. (In Russ.). DOI 10.21323/2414–438X‐2018–3–4–4–15 Томашевич И.Б. Белградский университет, Белград, Сербия Igor B. Tomasevic University of Belgrade, Belgrade, Serbia


Introduction
Color is primarily a personal experience.Without color, visual and emotional experience we have while looking at world around us, including our food is imperfect.Visual appearance is the first to arise for most foods.Color influences meat-purchasing decisions as well.For the consumers, major indicator of freshness and wholesomeness is discoloration, making color a major meat quality factor [1].This information is apparent to meat producers, retailers, and to researchers in meat science and technology, as well.Importance of color is also reflected in the fact that improving color stability of meat and meat products will influence their shelf life by increasing the time that meat is still visually acceptable to consumers at retail [2].
To ensure food conformity to consumer expectations, it is critical for the food processing industry to develop effective color inspection systems to measure the color information of food product.Traditionally, instrumental poultry meat color is assessed with a colorimeter [3].However, all colorimeters have the disadvantage that the surface to be measured must be uniform and rather small (~2-5 cm 2 ) [4] which influence bias in measurements.Another problem is that optically non-homogeneous medium such as poultry meat, refract, reflect, diffuse and absorb the light beam emitted by the colorimeter [5] causing deviations in all color dimensions evaluated.
With the aim of measuring food color rapidly and non-invasively, new objective and consistent methods are required for the effective color control of poultry meat.Among numerous new sensing technologies assessment of agricultural and food products, computer vision system (CVS) is a novel technology for food color evaluation [6,7,8].The aim of this review was to present the application of CVS for instrumental color evaluation of poultry meat, game meat and meat products with various physical properties and its advantages over the traditional color measuring method.

Samples of meat and meat products
The research was carried out on m. pectoralis major samples of three animals for each of the four poultry species (chicken, turkey, duck and goose) and five game meat species (quail, wild boar, rabbit, deer and pheasant).We selected the samples in a retail setting.Before color analysis, freshly cut meat samples, about 3.00 cm thick, were individually placed on white polystyrene foam trays with a consistent color and over wrapped with a transparent PVC film permeable to oxygen.Then they were placed in a bench refrigerator at 4 °C for 30 min to obtain myoglobin oxygenation.The PVC film was removed before color measurement.
Based on the treatment of raw materials and the individual processing steps and taking into account the processing technologies used, it is possible to classify processed meat products in six broad groups of processed meat products [9].In our research, within each product category, there were at least two and maximum four representative samples adding together 18 different meat products investigated.

Minolta CR-400 colorimeter
Minolta CR-400 colorimeter was used with 8 mm aperture, 2° observer, illuminant D65 and pulsed xenon lamp as a default light source.Glass cover was applied over the aperture port while measuring.A calibration of a device with white tile standard was performed before each analysis.

Computer vision system (CVS)
A Sony Alpha DSLR-A200 digital camera (10.2 Megapixel CCD sensor) was used.The camera was located vertically at a 30 cm distance from the sample (Figure 1).The camera setting was the following: shutter speed 1/6 s, manual operation mode, aperture Av F/11.0,ISO velocity Figure 1.Computer vision (image acquisition) system [10] 100, flash off, focal distance 30 mm, lens: DT-S18-70 mm f 3.5-5,6.Four Philips fluorescent lamps (Master Graphica TLD 965) with a color temperature of 6500 K were used for lighting the CVS.Each lamp was equipped with a designated light diffuser.In order to achieve the uniform light intensity on the sample, the lamps (60 cm length) were located at a 45° angle and 50 cm above the samples.Both the lamps and the camera were fixed inside a cubical (a = 80 cm) wooden box with a removable top (Figure 6).The box had an opening to the side for sample entry and the other on the top for visual inspection before and after the measurements.The internal walls of the box were coated with black opaque photographic cloth to diminish background light.
After the camera and the monitor were calibrated, as explained in the investigation of Tomasevic, Tomovic [10], the Adobe Photoshop CC (64 bit) software was used for image analysis.The colorimetric characteristics from RGB images were acquired using RAW photographs.They were measured on the digital image of the sample, using a Photoshop (31 x 31 pixels) Average Color Sampler Tool.

Color changes
Total color difference (ΔE) was determined by using the standard equation: Values for a C , b C , L C were obtained from the meat products using CVS, and for a M , b M , L M using Minolta.
Degree of difference of hue as the quantitative attribute of colorfulness chroma (C* ab ) was calculated according to Fernández-Vázquez, Stinco [11]: The difference in Chroma and lightness value was calculated using standard formulas: ) Hue difference ΔH was calculated according to Mokrzycki and Tatol [12]:

Similarity tests
The tests used were adopted from the investigation of Girolami, Napolitano [5] with slight modifications.For all the tests performed, 14 panelists were individually seated at a distance of approximately 60 cm from the calibrated monitor, equipped with a shade that reduces glare (Compushade Universal Monitor Hood, DulCO, USA), and from the meat samples presented inside the CVS wooden box.
For the test A, panelists were asked to individually analyze the color similarity between a digital image displayed on the monitor and a meat sample presented on polystyrene trays.They had up to 30 s to evaluate each sample by answering «yes» or «no».If yes, the panelists had the opportunity to rank the level of similarity according to a five-point Likert scale from 1 «very low», 2 «low», 3 «moderate», 4 «high» to 5 «very high».
Test B included displaying colors generated by Adobe Photoshop CC (2015) using the L*, a* and b* values obtained from both the CVS and Colorimeter (Minolta) data together on the monitor and panelists were asked to evaluate which of the two generated color chips was more similar to the sample of the product visualized on the monitor.
During the test C, the panelists were asked to evaluate the level of difference between the two color chips (colorimeter and CVS) displayed on the monitor ranking it according to a five-point Likert scale from 1 «very low», 2 «low», 3 «moderate», 4 «high» to 5 «very high».

Statistical analysis
The data gathered from the similarity tests (A, B) were analyzed to determine statistical significance based on the frequency of each response (χ2 One sample test), where the expected frequency was 50 %.In order to analyze data in respect to level of similarity (test A) and level of difference (Test C), one-way ANOVA was used.To distinguish statistical differences between the data, Tukey's post hoc tests were performed.

Poultry meat
The L*, a*, b*, chroma and hue angle values of poultry meat, measured with CVS and colorimeter in our experiment, were significantly different [13].The magnitude of colour difference between the two equipment's used is best represented by the total colour difference value (ΔE).The clear threshold for human meat-colour difference detec-

CVS Colorimeter
Chiken Turkey Duck Goose tion has not been established but a possible value could be around 2-6 [14].The values of ΔE in a range from 2-10 indicate that the difference in colour is perceptible at a glance and when they are larger than 10, we can conclude that colours are more opposite then similar [15].Therefore, with we ΔE = 18.5 for chicken meat an ΔE = 22.04 for turkey meat, we can concludethat the two systems measured their colour significantly different and even contrasting [13].Positive ΔL values for the same samples indicate that the colour measured with CVS was lighter than the colour obtained with colourimeter(Figure 2).However, the total color differences (ΔE) between the two methods, for duck and goose were half the values calculated for chicken and turkey.Yet, with the ΔE values above 10[13] these differences in color should be perceptible at a glance or considered more opposite then similar.Negative ΔL values for duck and goose breasts indicate that the color measured with CVS was darker than the color obtained with colorimeter (Figure 2).

Game meat
Game meat is darker red in appearance than meat from domestic animals [16], and is characterized by low L* values below 40, high a* values and low b* values which are indicative of the dark red color [17].However, the L*, a* and chroma values measured with CVS and colorimeter in our experiment were significantly different [18].Negative ΔL values for wild boar and deer meat indicate that the color measured with CVS was darker than the color obtained with colorimeter.All the a* values were higher when measured with CVS compared to colorimeter meaning that the color obtained with CVS was more «red» (or less «green») (Figure 3).The statistically significant differences between applied methods were not observed for b* and hue angle values.It is evident that differences in meat color and color stability between species can largely be attributed to differences in their activity, which influences the muscle fiber type, Mb concentration and intra muscular fat content of the meat, which in turn influences muscle color.Therefore, not all game meat is darker in color than meat from domestic animals [19].
The instrumental color values (L*, a*, b*, chroma and hue angle) obtained with CVS for lighter colored game meat samples (quail, pheasant and rabbit) were statistically different from the same values obtained with colorimeter [18].Positive ΔL values indicate that the color measured with CVS was lighter than the color obtained with colorimeter.All the a* values were much higher when measured with CVS compared to colorimeter meaning that the color obtained with CVS was more «red» (or less «green») (Figure 3).The positive difference in chroma (ΔC) meant that the CVS-generated color of quail and rabbit had greater intensity (were more saturated) then colorimeter-generated colors [18].The CVS-generated colors were in a clockwise direction from colorimeter-generated colors, representing a shift in the red direction (Figure 3), since all the Hue angle values were significantly higher when measured with colorimeter compared to CVS.The values of ΔE were in a range from 9.67-19.01indicating that for lighter colored game meat samples, the two systems measured their color significantly different [18]and in the case of rabbit meat even contrasting.

Meat products
When a color of uniformly-colored meat products was evaluated, the total color difference value (ΔE) ranged from 6.7 for Saveloy sausage up to 26.0 calculated for Pork prosciutto.For the majority of meat products with homogenous surfaces ΔE was around 10 [10].Positive ΔL values for uniformly-colored meat products indicate that the color measured with CVS was lighter than the color obtained with colorimeter.All the a* values were higher when measured with CVS compared to colorimeter meaning that the color obtained with CVS was more «red» (Figure 4) With the exception of pork prosciutto and raw sausage, all the b* measured with colorimeter were significantly higher than the values obtained with CVS [10], meaning that the colors of uniformly-colored meat products acquired with CVS were more «blue» (or less «yellow») compared to colorimeter-acquired color (Figure 4).The positive difference in chroma (ΔC) meant that the CVS color of cooked ham, pork and beef prosciutto and raw sausage, had greater intensity or were more saturated than colorimeter generated colors [10].The opposite was observed for the beef, chick-  [20] that CVS is a tool that can objectively specify color of cooked-hams.Bi-colored meat products, like mortadella, bacon, dry pork neck or pancetta, consisted of meat and fat segments that were larger than Minolta aperture size (8 mm) used in our experiment, allowing colorimeter to measure their color independently.The total color differences between the two methods of the meat segments were in a range from 7.3 up to 14.6 and for the fat parts in a range from 7.7 up to 12.9 [10].Meat segments were assessed in darker and fat segments in lighter colors when measured with CVS compared to colorimeter (Figure 5a.) Non-uniformly colored meat product was any product that has meat and fat parts that are too small (less than 8 mm) for colorimeter to independently assess their color.Therefore, when the color of beef and pork fermented sausage, and hamburger was measured, the L*, a*, b* colorimeter-generated values for both meat and fat parts were the same.Because CVS used 13 × 13 pixels Average Color Sampler Tool, it was capable of measuring the color of meat and fat parts independently.This resulted with the highest total meat-parts color difference (ΔE = 20.3)observed for beef fermented sausage, and maximum total fat-parts color difference (ΔE = 35.3)observed for pork fermented sausage [10].These extraordinary high values for total color differences [21]indicated that the colors assessed by the two methods were almost exact opposites [15].The color of meat parts measured with CVS was significantly darker, had greater intensity and were more saturated, compared to colorimeter-measured equivalents(Figure 5b).The opposite was observed for CVS-generated fat color.Due to the high variability and complex color distribution in non-uniformly colored meat products, the colorimeter was unable to assess accurately neither the color of meat nor the color of fat parts.Instead, colorimeter reproduced L*, a*, b* values that were somewhere «in between» the two segments.Our investigation is in concurrence with the conclusions of Girolami, Napolitano [22] that CVS is a tool that can objectively evaluate color of fermented sausages.

Similarity tests
The results of the first similarity test (test A) between the colour of the actual sample of meat products and the CVS obtained colour of the image displayed on the monitor, showed that the panelists found the digital images similar to the actual samples (P < 0.001).Frequency of similarity assessed by the panelists was 100 % for all poultry meat and game meat samples (Table 1).This means that 14 out of 14 panelist found that the actual colour of all samples was similar to the chip color generated with CVS.Frequency of similarity for meat products was also very high and ranged from 92.9 % for chicken pate, beef sausage, smoked bacon, dry pork neck and pancetta, to 100 % for all the other meat products samples.For poultry meat samples the level of similarity ranged from «low» to «moderate» and for game meat and meat products samples from «moderate» to «high».
Five-point Likert scale ranks from 1 «very low», 2 «low», 3 «moderate», 4 «high» to 5 «very high» Test B showed that the CVS-generated color chips were more similar to the sample of the poultry meat, game meat and meat products visualized on the monitor, compared to colorimeter-generated color chipsin all (100 %) individual trials performed (Table 1).
Test C regarding meat products revealed that, as assessed by the panelists, the magnitude of differences between the color chips generated by CVS and colorimeter and displayed on the monitor, ranged from 1.2 («very low») for Saveloy sausage to 4.2 («high») for Pork prosciutto.Highest level of difference between colors for poultry meat was observed in the case of turkey meat (4.7 -«very high») and for game meat samples with rabbit (4.2 -«high»).

Conclusions
We presume that one parameter influencing the difference among the meat and meat-products color measurements, between the two methods employed, could be the penetration depth of the illumination source.In our investigation, light employed in both devices had the same color temperature (6500 K) but the light interaction with a meat product samples was obviously device dependent.For the same reasons observed in meat color experiment of Girolami, Napolitano [5], we believe that the colorimeter could not be suitable for the color analysis of meat products.The reason is the translucent and optically non-homogenous matrix of the meat products due to the presence of different ingredients scattered inside it.The colorimeter is placed on the sample surface and the light penetration through meat product matrix must be higher than for CVS.This therefore causes multiple reflections and refractions where optical discontinuities are present, resulting in a diffusion of light (scattering) from the illuminations source [23], making the colorimeter measurements less accurate.Means in the same column with different small letters are significantly different (P < 0.05)

Figure 2 .
Figure2.Color of poultry meat as measured by the two methods[13]

Figure 3 .
Figure 3.Color of game meat as measured by the two methods[18]

Figure 4 .Figure 5 .
Figure 4. Color of uniformly colored meat products as measured by the two methods[10]

Table 1 .
Similarity tests results