- Email: [email protected]
Comparison of the predictive power of beef surface wavelet texture features at high and low magnification
Meat Science 82 (2009) 353–356
Contents lists available at ScienceDirect
Meat Science journal homepage: www.elsevier.com/locate/meatsci
Comparison of the predictive power of beef surface wavelet texture features at high and low magnification Patrick Jackman a,b, Da-Wen Sun a,*, Paul Allen b a b
FRCFT Research Group, University College Dublin, National University of Ireland, Agriculture and Food Science Centre, Belfield, Dublin 4, Ireland Ashtown Food Research Centre, Teagasc, Ashtown, Dublin 15, Ireland
a r t i c l e
i n f o
Article history: Received 9 May 2008 Received in revised form 22 January 2009 Accepted 9 February 2009
Keywords: Computer vision Image processing Beef Palatability Overall acceptability Genetic algorithms Magnification
a b s t r a c t Beef longissimus dorsi surface texture is an indicator used in predicting beef palatability by expert graders. Computer vision systems have previously used imaging at normal view to develop surface texture features with some success. Good models of beef overall acceptability using imaging at high magnification have been recently developed. As a comparison the same surface texture features were computed from the corresponding images at normal view and used to model overall acceptability. Both sets of texture features were also combined with muscle colour and marbling features and used to model overall acceptability. Models using texture features alone were more successful at normal modality. However colour and marbling features combined much better with texture features at high modality to yield the most accurate model of overall acceptability (r2 = 0.93). Accurate Partial Least Squares Regression (PLSR) models were computed at both modalities with and without inclusion of colour and marbling features. Addition of squared terms to the models failed to improve accuracy. Ó 2009 Elsevier Ltd. All rights reserved.
1. Introduction The surface texture properties of beef Longissimus Dorsi (LD) muscle are used by United States Department of Agriculture (USDA) expert graders as part of their palatability assessment after carcass quartering (USDA, 1997). Surface texture is a vitally important palatability indicator as it will reflect the amount of connective tissue in the muscle. Connective tissue is known to increase the toughness of beef (Li, Zhou, & Xu, 2008; Swatland, 2006). A fine texture indicates low connective tissue content while a coarse texture indicates high content (Li, Tan, & Shatadal, 2001). Surface texture will also reflect muscle fibre bundle size (Li et al., 2001). There are a number of essential problems with expert grading of surface texture properties. Firstly the judgement of expert graders suffers from subjectivity and inconsistency, secondly the muscle surface can only be examined unmagnified leading to some finer details being missed, thirdly the human grader has limited means of perceiving and analysing texture. A computerised vision system can avoid these problems and give better judgements of palatability based on surface texture. There are other palatability indicators used by USDA graders in forming their judgement such as LD colour and marbling, skeletal * Corresponding author. Tel.: +353 1 7167342; fax: +353 1 7167493. E-mail address: [email protected] (D.-W. Sun). URLs: http://www.ucd.ie/refrig, http://www.ucd.ie/sun (D.-W. Sun). 0309-1740/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.meatsci.2009.02.006
maturity and ribeye firmness (USDA, 1997). The same problems of subjectivity, inconsistency and limitation of perception and analysis arise with expert grading of these properties. Of these indicators LD colour and marbling are easily measured with any vision system that measures surface texture. Thus the computer vision system can easily combine surface texture features with LD colour and marbling when making a palatability judgement. These three indicators are linked to tenderness, juiciness and flavour (USMEF, 2007) which are the three most important aspects of palatability (Warriss, 2000). Previous studies with such computer vision systems have shown that surface texture properties of meat images can form a large part of a predictive model accounting for a substantial proportion of tenderness variability (Chandraratne, Samarasinghe, Kulasiri, & Bickerstaffe, 2006; Li, Tan, Martz, & Heymann, 1999; Tian, McCall, Dripps, Yu, & Gong, 2005). Typically meat surface texture has been expressed using classical algorithms such as pixel cooccurrence, difference histograms and run lengths (Chandraratne et al., 2006; Li et al., 1999; Tian et al., 2005). The advantage of perceiving texture in these ways is that they immediately make sense and can be easily understood by observation of an image. However analysis by Jackman, Sun, Du, Allen, and Downey (2008) and Huang et al. (1997) demonstrated that the wavelet transform is a superior means of expressing meat surface texture than classical algorithms. The wavelet transform imagines a texture greyscale image as a two dimensional wave of limited duration with the pixel grey
354
P. Jackman et al. / Meat Science 82 (2009) 353–356
level being the height of the wave at that point. The shape of the wave can broken down into fundamental elements called wavelets which when added together recreate the wave. Using a small subset of these wavelets a very close approximation of the original wave is possible. One such subset is dyadic scales (scales of 2k). Thus the wavelet transform can compress image data into a relatively small number of features without substantial loss of information. Hence wavelets can provide highly efficient texture analysis. The mathematical foundations of wavelets are given by Kaiser (1994). The benefits of applying a wavelet transform to food quality inspection data are discussed in detail by Singh, Choudhary, Jayas, and Paliwal (2008). The Federal Bureau of Investigation (FBI) in the USA applies the wavelet transform to compress fingerprint images without substantial decline in fingerprint image quality. Wickerhauser (1994) explains this in some more detail. Previous studies used imaging at normal view to generate the surface texture features. It was proposed by Li et al. (2001) that more useful surface texture features could be found at higher magnifications as this would allow a more detailed view of the muscle fibre bundles. On the basis of this proposal surface texture features of images at high magnification rather than normal view were used by Jackman et al. (2008) to develop predictive models of beef palatability without and in combination with L. Dorsi (LD) colour and marbling features. Models produced using the surface texture features at high magnification proved very successful at modelling important palatability measurements. Jackman et al. (2008) used a corresponding image at normal view to evaluate LD colour and marbling features for each sample. This image can also be used for surface texture analysis as described by Jackman, Sun, Du, and Allen (2009) in similar work where high magnification images were not available for all samples. Hence by extracting the same surface texture features from these corresponding normal view images as were extracted from the high magnification images a comparison can be made of the predictive power of surface texture features at normal view and high magnification. Some of the previous studies (Jackman et al., 2008, 2009; Li et al., 1999; Tian et al., 2005) have shown how surface texture features can combine effectively with LD colour and marbling in modelling palatability. Hence the predictive power of both sets of texture features should also be evaluated in terms of how well they combine with LD colour and marbling features.
magnification. High magnification images of 5 spots on the LD muscle were taken. Image acquisition was in RGB compressed TIFF form (Tag Image File Format). Images were decompressed into BMP form (Bitmap) with conversion software Optimas (Meyer Instruments, Houston, TX, USA, www.meyerinst.com) before analysis in Matlab (Mathworks, Natick, MA, USA, www.mathworks.com). A greyscale was used to reveal surface texture. The greyscale chosen was colour saturation as this was previously found to be highly effective for beef by Li et al. (2001). To create this greyscale each image was transformed into the HSI colour space, with the second channel in the image being the colour saturation. Each image at high magnification was trimmed into 512 512 size for wavelet analysis. The normal view images could not be trimmed to 512 512 size as this would go beyond the edges of the muscle. Thus the nearest dyadic scale was used. Hence each image at normal view had a window of 256 256 size on the centre of the LD muscle extracted. Cropping the high magnification images to 256 256 size was not performed as this would reduce the area of muscle examined possibly reducing the representivity of the data. Fig. 1 shows a saturation image at normal view and Fig. 2 illustrates a saturation image at high magnification from the same sample. The image acquisition system failed for 2 samples thus reducing the sample number to 46. 2.3. Sensory property evaluation Sensory panel assessment was carried out on the 14-day aged steaks in the same manner as described by Jackman et al. (2009). A steak was cooked to 71 °C internal temperature in an electric grill (HB 90420, Siemens-Electerogerate, GmbH, Munich, Germany). Square pieces 12.5 mm in dimension were cut from the steak and served to eight in house panellist’s who evaluated overall acceptability. The sum of acceptability scores for each sample was used for data analysis. 2.4. Data processing Partial least squares regression (PLSR) predictive models were generated for overall acceptability with Unscrambler (Camo Software, Woodbridge, NJ, USA, www.camo.com). The models were validated by full cross validation. The models used the Symlet (Symmetric modification of the Daubechie wavelet) to express sur-
2. Materials and methods 2.1. Sample preparation Thirty two heifers approximately two years old from a pilot scale abattoir (Meat Industrial Development Unit, Teagasc Ashtown Food Research Centre, Dublin) were prepared and slaughtered in the same way as described by Du, Sun, Jackman, and Allen (2008). Sixteen carcasses were then chilled at 2 °C for 2 days. One side of the remaining 16 were then chilled at 5 °C for 2 days and the other at 5 °C also for 2 days. This effectively produced 48 samples. A steak for computer vision analysis and a duplicate for sensory panel assessment were excised from each sample. The steaks for sensory panel assessment were vacuum packed and aged for a further 12 days, after which they were transferred to a 20 °C freezer for storage. 2.2. Image acquisition and processing The acquisition system and procedure for imaging at normal view is the same as described by Jackman et al. (2009). For imaging at high magnification a telescope (CFM2, Infinivar, Boulder, CO, USA) was attached to the camera. This telescope offers 6.4 times Fig. 1. A saturation channel image at normal view.
355
P. Jackman et al. / Meat Science 82 (2009) 353–356 Table 2 Genetic algorithm settings. Population = 50 100 Generations Uniform creation Rank fitness Stochastic uniform selection 10% elitism 0.7 crossover fraction Gaussian mutation, scale = 1, shrink = 1 Scattered crossover Forward migration, fraction = 0.2, interval = 20 Max condition number estimate XXT = 1,000,000 No hybrid function Fitness function not vectorised
and overall marbling fleck number density were calculated from the marbling image. The best PLSR subsets were selected with the same genetic algorithms. 3. Results and discussion 3.1. PLSR modelling Fig. 2. A saturation channel image from the same sample at high magnification.
face texture. Jackman et al. (2008) identified this as the most effective means of expressing surface texture. Each 512 512 image yielded 37 root mean square values for each order of the Symlet decomposition used (2nd, 3rd, 4th, 5th, 6th, 7th and 8th order). Similarly each 256 256 image yielded 33 root mean square values for each order of the Symlet decomposition used. Hence 259 values per 512 512 image and 231 values per 256 256 image were computed. The values were then standardised. Best linear PLSR subsets for overall acceptability were selected by a genetic algorithm. The algorithm parameters are given in Table 1. Another genetic algorithm then searched for useful squared terms considering only terms present in the linear models, whose algorithm parameters are given in Table 2. Searching for useful more squared terms is much easier as the number of possibilities is greatly reduced by the fact that only squared terms based on the linear model terms are allowable. Hence a much simpler and faster genetic algorithm was used. Corresponding models were computed including LD colour and marbling features. The images at normal view were segmented with the method of Jackman et al. (2009) into an LD muscle image and a marbling image. The mean, standard deviation, skewness, kurtosis and interquartile range of red, green and blue were calculated from the LD muscle image. The same five histogram features of marbling fleck size as well as overall marbling fleck area density
Table 1 Genetic algorithm parameters. Population = 1000 100 Generations Customised creation Rank fitness Stochastic uniform selection 1% elitism 0.7 crossover fraction Gaussian mutation, scale = 1, shrink = 1 Scattered crossover Forward migration, fraction = 0.2, interval = 20 Max condition number estimate XXT = 1,000,000 No hybrid function Fitness function not vectorised
The results of PLSR regression models are presented in Table 3 as r2 values, showing the proportion of variance accounted for by the model, and as the model root mean square prediction error relative to the standard deviation of the sensory measurement being predicted (RRR). Linear and squared terms were included to account for possible non-linearity as suggested by Huang et al. (1997). All models were subjected to tests of legitimacy: a well behaved residual curve, a stable regression matrix and a failure to predict a random variable. It can be seen that surface texture features extracted from the images at normal view contained more useful information on beef palatability. However in contrast when colour and marbling features are incorporated the normal view model improved only slightly (r2 = 0.03) while the high magnification model improved more substantially (r2 = 0.14). This difference can be explained by the fact that muscle colour is a proxy for the muscle pH, which affects the electric charge on the muscle proteins and hence how tightly the muscle fibres are packed due to the protein’s stronger association with water in the muscle if ultimate pH is high (Abril et al., 2001). This will be reflected in the muscle macrotexture. Thus it can be expected that colour and normal view surface texture feature will contain some similar information. In contrast at high magnification more of the muscles microtexture will be observed which can be expected to contain less similar information to colour. Accurate modelling is possible in all 4 cases as an r2 of greater than 0.8 is achieved. Shiranita, Hayashi, Otsubo, Miyajima, and Takiyama (2000) set this benchmark for an accurate multiple regression model. Only the high magnification model incorporating colour and marbling features met another important benchmark of a RRR of less than one third. Addition of squared terms
Table 3 Best PLSR models. Texture with colour & marbling
r2
RRR
Factors
High magnification Normal view
0.93 0.87
0.26 0.36
18 12
Texture only
r2
RRR
Factors
High magnification Normal view
0.79 0.84
0.47 0.40
14 17
356
P. Jackman et al. / Meat Science 82 (2009) 353–356
of non-linearity was found. Other palatability indicators such as skeletal maturity may provide further information.
Table 4 Regression variables. Colour, marbling & texture high
Colour, marbling & texture normal
Texture only high
Texture only normal
h7-2 d5-2 h5-3 h9-3 v5-3 d7-3 ap0-3 v8-4 v6-5 v7-5 d5-5 d8-6 h5-7 v5-7 d7-7 d8-7 h4-8 v3-8 v7-8 Mean red Kurt green Kurt blue Stdev Fat
h4-2 v2-4 v3-4 v6-4 v7-4 v8-4 d7-4 v1-5 v5-6 v6-6 v8-6 v2-7 v7-7 h4-8 d7-8 Mean blue
d6-2 d8-2 v3-3 v7-3 d5-3 d9-3 v7-4 h9-5 d6-5 d8-5 ap7-5 h4-6 d6-6 d9-6 h5-7 v5-7 v6-7 d5-7
h4-2 h8-2 v6-2 v3-3 d4-3 h8-4 v3-4 h4-5 h5-5 v4-5 h7-6 d7-6 v7-7 d4-7 d7-7 ap7-7 v5-8 d3-8
Note: For texture feature xm-n, ap = approximation, d = diagonal, h = horizontal, v = vertical, m = level of decomposition and n = symlet order.
proved ineffective in all 4 cases. The results show that the proposal of Li et al. (2001) that more useful texture features could be found at higher magnification appears to be justified. However this is only true when LD colour and marbling features are included. This indicates that the texture features at normal view do contain similar information to colour. This can be seen in Table 4 which gives the regression variables where only a single colour variable is included in the best subset at normal view while 3 are included at high magnification. Furthermore removal of the single variable did not detract substantially from the accuracy of the normal view model. 3.2. Future work The accurate models of acceptability show that beef palatability can be modelled by computer vision with the information contained in LD colour, marbling and surface texture features. This is a vindication of the proposal that image analysis could replace the human expert grader. Further work may provide better modelling. Three dimensional imaging could provide more detailed texture features and would at least avoid the problem of surface damage caused by the cutting of the LD muscle to expose the ribeye surface. The use of artificial intelligence to enhance predictive models of meat palatability has previously proven successful (Chandraratne et al., 2006; Li et al., 1999; Tian et al., 2005). The current models might be improved with neural networks or fuzzy inference systems built on the best subsets although little evidence
4. Conclusions Results show that when surface texture features alone are used to predict the palatability of grass fed beef the texture features at normal view contain stronger predictive power. However LD colour and marbling features combine much better with texture features at high magnification and produce more accurate modelling. Thus high magnification imaging can be considered more useful. Accurate models of acceptability can be computed in all cases. Acknowledgements This work is part of a FIRM project administered and partfunded by the Department of Agriculture, Fisheries and Food. Patrick Jackman is a Teagasc Walsh Fellow. References Abril, M., Campo, M. M., Onenc, A., Sanudo, C., Alberti, P., & Negueruela, A. I. (2001). Beef colour evolution as a function of ultimate pH. Meat Science, 58(1), 69–78. Chandraratne, M. R., Samarasinghe, S., Kulasiri, D., & Bickerstaffe, R. (2006). Prediction of lamb tenderness using image surface texture features. Journal of Food Engineering, 77(3), 492–499. Du, C.-J., Sun, D.-W., Jackman, P., & Allen, P. (2008). Development of a hybrid image processing algorithm for automatic evaluation of intramuscular fat content in beef M. longissimus dorsi. Meat Science, 80(4), 1231–1237. Huang, Y., Lacey, R. E., Moore, L. L., Miller, R. K., Whittaker, A. D., & Ophir, J. (1997). Wavelet textural features from ultrasonic elastograms for meat quality prediction. Transactions of the ASAE, 40(6), 1741–1748. Jackman, P., Sun, D.-W., Du, C.-J., Allen, P., & Downey, G. (2008). Prediction of beef eating quality from colour, marbling and wavelet texture features. Meat Science, 80(4), 1273–1281. Jackman, P., Sun, D.-W., Du, C.-J., & Allen, P. (2009). Prediction of beef eating qualities from colour, marbling and wavelet surface texture features using homogeneous carcass treatment. Pattern Recognition, 42(5), 751–763. Kaiser, G. (1994). A friendly guide to wavelets. Cambridge, MA, USA: Birkhauser Boston. Li, J., Tan, J., & Shatadal, P. (2001). Classification of tough and tender beef by image texture analysis. Meat Science, 57(4), 341–346. Li, J., Tan, J., Martz, F. A., & Heymann, H. (1999). Image texture features as indicators of beef tenderness. Meat Science, 53(1), 17–22. Li, C. B., Zhou, G. H., & Xu, X. L. (2008). Dynamical changes of beef intramuscular connective tissue and muscle fiber during heating and their effects on beef shear force. Food and Bioprocess Technology, doi:10.1007/s11947-008-0117-3. Shiranita, K., Hayashi, K., Otsubo, A., Miyajima, T., & Takiyama, R. (2000). Grading meat quality by image processing. Pattern Recognition, 33(1), 97–104. Singh, C. B., Choudhary, R., Jayas, D. S., & Paliwal, J. (2008). Wavelet analysis of signals in agriculture and food quality inspection. Food and Bioprocess Technology, doi:10.1007/s11947-008-0093-7. Swatland, H. J. (2006). Stratification of connective tissue toughness in beef roasts assessed by simultaneous fluorometry and penetrometry. Food Research International, 39, 1106–1109. Tian, Y. Q., McCall, D. G., Dripps, W., Yu, Q., & Gong, P. (2005). Using computer vision technology to evaluate the meat tenderness of grazing beef. Food Australia, 57(8), 322–326. USMEF (2007). Backgrounder grading – trade library november, Denver, USA. USDA (1997). United States standards for graders of carcass beef. Washington, DC: USDA. Warriss, P. D. (2000). Meat science. An introductory text. Wallingford: CABI Publishing. Wickerhauser, M. V. (1994). Adapted wavelet analysis from theory to software algorithms. A.K. Peters.
Contents lists available at ScienceDirect
Meat Science journal homepage: www.elsevier.com/locate/meatsci
Comparison of the predictive power of beef surface wavelet texture features at high and low magnification Patrick Jackman a,b, Da-Wen Sun a,*, Paul Allen b a b
FRCFT Research Group, University College Dublin, National University of Ireland, Agriculture and Food Science Centre, Belfield, Dublin 4, Ireland Ashtown Food Research Centre, Teagasc, Ashtown, Dublin 15, Ireland
a r t i c l e
i n f o
Article history: Received 9 May 2008 Received in revised form 22 January 2009 Accepted 9 February 2009
Keywords: Computer vision Image processing Beef Palatability Overall acceptability Genetic algorithms Magnification
a b s t r a c t Beef longissimus dorsi surface texture is an indicator used in predicting beef palatability by expert graders. Computer vision systems have previously used imaging at normal view to develop surface texture features with some success. Good models of beef overall acceptability using imaging at high magnification have been recently developed. As a comparison the same surface texture features were computed from the corresponding images at normal view and used to model overall acceptability. Both sets of texture features were also combined with muscle colour and marbling features and used to model overall acceptability. Models using texture features alone were more successful at normal modality. However colour and marbling features combined much better with texture features at high modality to yield the most accurate model of overall acceptability (r2 = 0.93). Accurate Partial Least Squares Regression (PLSR) models were computed at both modalities with and without inclusion of colour and marbling features. Addition of squared terms to the models failed to improve accuracy. Ó 2009 Elsevier Ltd. All rights reserved.
1. Introduction The surface texture properties of beef Longissimus Dorsi (LD) muscle are used by United States Department of Agriculture (USDA) expert graders as part of their palatability assessment after carcass quartering (USDA, 1997). Surface texture is a vitally important palatability indicator as it will reflect the amount of connective tissue in the muscle. Connective tissue is known to increase the toughness of beef (Li, Zhou, & Xu, 2008; Swatland, 2006). A fine texture indicates low connective tissue content while a coarse texture indicates high content (Li, Tan, & Shatadal, 2001). Surface texture will also reflect muscle fibre bundle size (Li et al., 2001). There are a number of essential problems with expert grading of surface texture properties. Firstly the judgement of expert graders suffers from subjectivity and inconsistency, secondly the muscle surface can only be examined unmagnified leading to some finer details being missed, thirdly the human grader has limited means of perceiving and analysing texture. A computerised vision system can avoid these problems and give better judgements of palatability based on surface texture. There are other palatability indicators used by USDA graders in forming their judgement such as LD colour and marbling, skeletal * Corresponding author. Tel.: +353 1 7167342; fax: +353 1 7167493. E-mail address: [email protected] (D.-W. Sun). URLs: http://www.ucd.ie/refrig, http://www.ucd.ie/sun (D.-W. Sun). 0309-1740/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.meatsci.2009.02.006
maturity and ribeye firmness (USDA, 1997). The same problems of subjectivity, inconsistency and limitation of perception and analysis arise with expert grading of these properties. Of these indicators LD colour and marbling are easily measured with any vision system that measures surface texture. Thus the computer vision system can easily combine surface texture features with LD colour and marbling when making a palatability judgement. These three indicators are linked to tenderness, juiciness and flavour (USMEF, 2007) which are the three most important aspects of palatability (Warriss, 2000). Previous studies with such computer vision systems have shown that surface texture properties of meat images can form a large part of a predictive model accounting for a substantial proportion of tenderness variability (Chandraratne, Samarasinghe, Kulasiri, & Bickerstaffe, 2006; Li, Tan, Martz, & Heymann, 1999; Tian, McCall, Dripps, Yu, & Gong, 2005). Typically meat surface texture has been expressed using classical algorithms such as pixel cooccurrence, difference histograms and run lengths (Chandraratne et al., 2006; Li et al., 1999; Tian et al., 2005). The advantage of perceiving texture in these ways is that they immediately make sense and can be easily understood by observation of an image. However analysis by Jackman, Sun, Du, Allen, and Downey (2008) and Huang et al. (1997) demonstrated that the wavelet transform is a superior means of expressing meat surface texture than classical algorithms. The wavelet transform imagines a texture greyscale image as a two dimensional wave of limited duration with the pixel grey
354
P. Jackman et al. / Meat Science 82 (2009) 353–356
level being the height of the wave at that point. The shape of the wave can broken down into fundamental elements called wavelets which when added together recreate the wave. Using a small subset of these wavelets a very close approximation of the original wave is possible. One such subset is dyadic scales (scales of 2k). Thus the wavelet transform can compress image data into a relatively small number of features without substantial loss of information. Hence wavelets can provide highly efficient texture analysis. The mathematical foundations of wavelets are given by Kaiser (1994). The benefits of applying a wavelet transform to food quality inspection data are discussed in detail by Singh, Choudhary, Jayas, and Paliwal (2008). The Federal Bureau of Investigation (FBI) in the USA applies the wavelet transform to compress fingerprint images without substantial decline in fingerprint image quality. Wickerhauser (1994) explains this in some more detail. Previous studies used imaging at normal view to generate the surface texture features. It was proposed by Li et al. (2001) that more useful surface texture features could be found at higher magnifications as this would allow a more detailed view of the muscle fibre bundles. On the basis of this proposal surface texture features of images at high magnification rather than normal view were used by Jackman et al. (2008) to develop predictive models of beef palatability without and in combination with L. Dorsi (LD) colour and marbling features. Models produced using the surface texture features at high magnification proved very successful at modelling important palatability measurements. Jackman et al. (2008) used a corresponding image at normal view to evaluate LD colour and marbling features for each sample. This image can also be used for surface texture analysis as described by Jackman, Sun, Du, and Allen (2009) in similar work where high magnification images were not available for all samples. Hence by extracting the same surface texture features from these corresponding normal view images as were extracted from the high magnification images a comparison can be made of the predictive power of surface texture features at normal view and high magnification. Some of the previous studies (Jackman et al., 2008, 2009; Li et al., 1999; Tian et al., 2005) have shown how surface texture features can combine effectively with LD colour and marbling in modelling palatability. Hence the predictive power of both sets of texture features should also be evaluated in terms of how well they combine with LD colour and marbling features.
magnification. High magnification images of 5 spots on the LD muscle were taken. Image acquisition was in RGB compressed TIFF form (Tag Image File Format). Images were decompressed into BMP form (Bitmap) with conversion software Optimas (Meyer Instruments, Houston, TX, USA, www.meyerinst.com) before analysis in Matlab (Mathworks, Natick, MA, USA, www.mathworks.com). A greyscale was used to reveal surface texture. The greyscale chosen was colour saturation as this was previously found to be highly effective for beef by Li et al. (2001). To create this greyscale each image was transformed into the HSI colour space, with the second channel in the image being the colour saturation. Each image at high magnification was trimmed into 512 512 size for wavelet analysis. The normal view images could not be trimmed to 512 512 size as this would go beyond the edges of the muscle. Thus the nearest dyadic scale was used. Hence each image at normal view had a window of 256 256 size on the centre of the LD muscle extracted. Cropping the high magnification images to 256 256 size was not performed as this would reduce the area of muscle examined possibly reducing the representivity of the data. Fig. 1 shows a saturation image at normal view and Fig. 2 illustrates a saturation image at high magnification from the same sample. The image acquisition system failed for 2 samples thus reducing the sample number to 46. 2.3. Sensory property evaluation Sensory panel assessment was carried out on the 14-day aged steaks in the same manner as described by Jackman et al. (2009). A steak was cooked to 71 °C internal temperature in an electric grill (HB 90420, Siemens-Electerogerate, GmbH, Munich, Germany). Square pieces 12.5 mm in dimension were cut from the steak and served to eight in house panellist’s who evaluated overall acceptability. The sum of acceptability scores for each sample was used for data analysis. 2.4. Data processing Partial least squares regression (PLSR) predictive models were generated for overall acceptability with Unscrambler (Camo Software, Woodbridge, NJ, USA, www.camo.com). The models were validated by full cross validation. The models used the Symlet (Symmetric modification of the Daubechie wavelet) to express sur-
2. Materials and methods 2.1. Sample preparation Thirty two heifers approximately two years old from a pilot scale abattoir (Meat Industrial Development Unit, Teagasc Ashtown Food Research Centre, Dublin) were prepared and slaughtered in the same way as described by Du, Sun, Jackman, and Allen (2008). Sixteen carcasses were then chilled at 2 °C for 2 days. One side of the remaining 16 were then chilled at 5 °C for 2 days and the other at 5 °C also for 2 days. This effectively produced 48 samples. A steak for computer vision analysis and a duplicate for sensory panel assessment were excised from each sample. The steaks for sensory panel assessment were vacuum packed and aged for a further 12 days, after which they were transferred to a 20 °C freezer for storage. 2.2. Image acquisition and processing The acquisition system and procedure for imaging at normal view is the same as described by Jackman et al. (2009). For imaging at high magnification a telescope (CFM2, Infinivar, Boulder, CO, USA) was attached to the camera. This telescope offers 6.4 times Fig. 1. A saturation channel image at normal view.
355
P. Jackman et al. / Meat Science 82 (2009) 353–356 Table 2 Genetic algorithm settings. Population = 50 100 Generations Uniform creation Rank fitness Stochastic uniform selection 10% elitism 0.7 crossover fraction Gaussian mutation, scale = 1, shrink = 1 Scattered crossover Forward migration, fraction = 0.2, interval = 20 Max condition number estimate XXT = 1,000,000 No hybrid function Fitness function not vectorised
and overall marbling fleck number density were calculated from the marbling image. The best PLSR subsets were selected with the same genetic algorithms. 3. Results and discussion 3.1. PLSR modelling Fig. 2. A saturation channel image from the same sample at high magnification.
face texture. Jackman et al. (2008) identified this as the most effective means of expressing surface texture. Each 512 512 image yielded 37 root mean square values for each order of the Symlet decomposition used (2nd, 3rd, 4th, 5th, 6th, 7th and 8th order). Similarly each 256 256 image yielded 33 root mean square values for each order of the Symlet decomposition used. Hence 259 values per 512 512 image and 231 values per 256 256 image were computed. The values were then standardised. Best linear PLSR subsets for overall acceptability were selected by a genetic algorithm. The algorithm parameters are given in Table 1. Another genetic algorithm then searched for useful squared terms considering only terms present in the linear models, whose algorithm parameters are given in Table 2. Searching for useful more squared terms is much easier as the number of possibilities is greatly reduced by the fact that only squared terms based on the linear model terms are allowable. Hence a much simpler and faster genetic algorithm was used. Corresponding models were computed including LD colour and marbling features. The images at normal view were segmented with the method of Jackman et al. (2009) into an LD muscle image and a marbling image. The mean, standard deviation, skewness, kurtosis and interquartile range of red, green and blue were calculated from the LD muscle image. The same five histogram features of marbling fleck size as well as overall marbling fleck area density
Table 1 Genetic algorithm parameters. Population = 1000 100 Generations Customised creation Rank fitness Stochastic uniform selection 1% elitism 0.7 crossover fraction Gaussian mutation, scale = 1, shrink = 1 Scattered crossover Forward migration, fraction = 0.2, interval = 20 Max condition number estimate XXT = 1,000,000 No hybrid function Fitness function not vectorised
The results of PLSR regression models are presented in Table 3 as r2 values, showing the proportion of variance accounted for by the model, and as the model root mean square prediction error relative to the standard deviation of the sensory measurement being predicted (RRR). Linear and squared terms were included to account for possible non-linearity as suggested by Huang et al. (1997). All models were subjected to tests of legitimacy: a well behaved residual curve, a stable regression matrix and a failure to predict a random variable. It can be seen that surface texture features extracted from the images at normal view contained more useful information on beef palatability. However in contrast when colour and marbling features are incorporated the normal view model improved only slightly (r2 = 0.03) while the high magnification model improved more substantially (r2 = 0.14). This difference can be explained by the fact that muscle colour is a proxy for the muscle pH, which affects the electric charge on the muscle proteins and hence how tightly the muscle fibres are packed due to the protein’s stronger association with water in the muscle if ultimate pH is high (Abril et al., 2001). This will be reflected in the muscle macrotexture. Thus it can be expected that colour and normal view surface texture feature will contain some similar information. In contrast at high magnification more of the muscles microtexture will be observed which can be expected to contain less similar information to colour. Accurate modelling is possible in all 4 cases as an r2 of greater than 0.8 is achieved. Shiranita, Hayashi, Otsubo, Miyajima, and Takiyama (2000) set this benchmark for an accurate multiple regression model. Only the high magnification model incorporating colour and marbling features met another important benchmark of a RRR of less than one third. Addition of squared terms
Table 3 Best PLSR models. Texture with colour & marbling
r2
RRR
Factors
High magnification Normal view
0.93 0.87
0.26 0.36
18 12
Texture only
r2
RRR
Factors
High magnification Normal view
0.79 0.84
0.47 0.40
14 17
356
P. Jackman et al. / Meat Science 82 (2009) 353–356
of non-linearity was found. Other palatability indicators such as skeletal maturity may provide further information.
Table 4 Regression variables. Colour, marbling & texture high
Colour, marbling & texture normal
Texture only high
Texture only normal
h7-2 d5-2 h5-3 h9-3 v5-3 d7-3 ap0-3 v8-4 v6-5 v7-5 d5-5 d8-6 h5-7 v5-7 d7-7 d8-7 h4-8 v3-8 v7-8 Mean red Kurt green Kurt blue Stdev Fat
h4-2 v2-4 v3-4 v6-4 v7-4 v8-4 d7-4 v1-5 v5-6 v6-6 v8-6 v2-7 v7-7 h4-8 d7-8 Mean blue
d6-2 d8-2 v3-3 v7-3 d5-3 d9-3 v7-4 h9-5 d6-5 d8-5 ap7-5 h4-6 d6-6 d9-6 h5-7 v5-7 v6-7 d5-7
h4-2 h8-2 v6-2 v3-3 d4-3 h8-4 v3-4 h4-5 h5-5 v4-5 h7-6 d7-6 v7-7 d4-7 d7-7 ap7-7 v5-8 d3-8
Note: For texture feature xm-n, ap = approximation, d = diagonal, h = horizontal, v = vertical, m = level of decomposition and n = symlet order.
proved ineffective in all 4 cases. The results show that the proposal of Li et al. (2001) that more useful texture features could be found at higher magnification appears to be justified. However this is only true when LD colour and marbling features are included. This indicates that the texture features at normal view do contain similar information to colour. This can be seen in Table 4 which gives the regression variables where only a single colour variable is included in the best subset at normal view while 3 are included at high magnification. Furthermore removal of the single variable did not detract substantially from the accuracy of the normal view model. 3.2. Future work The accurate models of acceptability show that beef palatability can be modelled by computer vision with the information contained in LD colour, marbling and surface texture features. This is a vindication of the proposal that image analysis could replace the human expert grader. Further work may provide better modelling. Three dimensional imaging could provide more detailed texture features and would at least avoid the problem of surface damage caused by the cutting of the LD muscle to expose the ribeye surface. The use of artificial intelligence to enhance predictive models of meat palatability has previously proven successful (Chandraratne et al., 2006; Li et al., 1999; Tian et al., 2005). The current models might be improved with neural networks or fuzzy inference systems built on the best subsets although little evidence
4. Conclusions Results show that when surface texture features alone are used to predict the palatability of grass fed beef the texture features at normal view contain stronger predictive power. However LD colour and marbling features combine much better with texture features at high magnification and produce more accurate modelling. Thus high magnification imaging can be considered more useful. Accurate models of acceptability can be computed in all cases. Acknowledgements This work is part of a FIRM project administered and partfunded by the Department of Agriculture, Fisheries and Food. Patrick Jackman is a Teagasc Walsh Fellow. References Abril, M., Campo, M. M., Onenc, A., Sanudo, C., Alberti, P., & Negueruela, A. I. (2001). Beef colour evolution as a function of ultimate pH. Meat Science, 58(1), 69–78. Chandraratne, M. R., Samarasinghe, S., Kulasiri, D., & Bickerstaffe, R. (2006). Prediction of lamb tenderness using image surface texture features. Journal of Food Engineering, 77(3), 492–499. Du, C.-J., Sun, D.-W., Jackman, P., & Allen, P. (2008). Development of a hybrid image processing algorithm for automatic evaluation of intramuscular fat content in beef M. longissimus dorsi. Meat Science, 80(4), 1231–1237. Huang, Y., Lacey, R. E., Moore, L. L., Miller, R. K., Whittaker, A. D., & Ophir, J. (1997). Wavelet textural features from ultrasonic elastograms for meat quality prediction. Transactions of the ASAE, 40(6), 1741–1748. Jackman, P., Sun, D.-W., Du, C.-J., Allen, P., & Downey, G. (2008). Prediction of beef eating quality from colour, marbling and wavelet texture features. Meat Science, 80(4), 1273–1281. Jackman, P., Sun, D.-W., Du, C.-J., & Allen, P. (2009). Prediction of beef eating qualities from colour, marbling and wavelet surface texture features using homogeneous carcass treatment. Pattern Recognition, 42(5), 751–763. Kaiser, G. (1994). A friendly guide to wavelets. Cambridge, MA, USA: Birkhauser Boston. Li, J., Tan, J., & Shatadal, P. (2001). Classification of tough and tender beef by image texture analysis. Meat Science, 57(4), 341–346. Li, J., Tan, J., Martz, F. A., & Heymann, H. (1999). Image texture features as indicators of beef tenderness. Meat Science, 53(1), 17–22. Li, C. B., Zhou, G. H., & Xu, X. L. (2008). Dynamical changes of beef intramuscular connective tissue and muscle fiber during heating and their effects on beef shear force. Food and Bioprocess Technology, doi:10.1007/s11947-008-0117-3. Shiranita, K., Hayashi, K., Otsubo, A., Miyajima, T., & Takiyama, R. (2000). Grading meat quality by image processing. Pattern Recognition, 33(1), 97–104. Singh, C. B., Choudhary, R., Jayas, D. S., & Paliwal, J. (2008). Wavelet analysis of signals in agriculture and food quality inspection. Food and Bioprocess Technology, doi:10.1007/s11947-008-0093-7. Swatland, H. J. (2006). Stratification of connective tissue toughness in beef roasts assessed by simultaneous fluorometry and penetrometry. Food Research International, 39, 1106–1109. Tian, Y. Q., McCall, D. G., Dripps, W., Yu, Q., & Gong, P. (2005). Using computer vision technology to evaluate the meat tenderness of grazing beef. Food Australia, 57(8), 322–326. USMEF (2007). Backgrounder grading – trade library november, Denver, USA. USDA (1997). United States standards for graders of carcass beef. Washington, DC: USDA. Warriss, P. D. (2000). Meat science. An introductory text. Wallingford: CABI Publishing. Wickerhauser, M. V. (1994). Adapted wavelet analysis from theory to software algorithms. A.K. Peters.