Color (gray level) estimation during offee roasting Proeedings of European Congress of Chemial Engineering (ECCE-6) Copenhagen, 16-20 September 2007 Color (gray-level) estimation during offee roasting J. A. Hernández a*, B. Heyd b, C. Irles, G. Trystram b. a Researh Center of Engineering and Applied Sienes (CIICAp), Autonomous University of Morelos State (UAEM); Av. Universidad No. 1001 Col. Chamilpa, C.P. 62210, Cuernavaa, Morelos, Mexio. b Joint Researh Unit Food Proess Engineering (Cemagref, ENSIA, INAPG, INRA) ENSIA, 1 avenue des Olympiades, 91744 Massy Cedex Frane. National Institute of Perinatology, Montes Urales 800, Lomas de Virreyes, C.P. 11000, Méxio D.F. Abstrat In order to optimize the quality of roasted offee, it is important to measure and to ontrol a large number of fators during the proess. Image analysis enables the on-line measurement of essential values suh as bean olor and surfae area. However, it is diffiult to apply this tehnique to offee roasting. In this industry, olor is a key variable whih determines the quality of the final produt, but it is evaluated out-line by the roast master. By this raison, it is neessary to developer a tehnique to olor and surfae estimate. Consequently, this work propose a method to determine olor and surfae area using images analysis and a mathematial model based in artifiial neural network for estimate the olor (gray values) during roasting offee. The mathematial model onsider as input variable the time and the temperature of the beans. A feedforward networks with one hidden layer is used to predit the gray values. For the network, the Levenberg-Marquardt learning algorithm, the hyperboli tangent sigmoid transfer-funtion and the linear transfer-funtion were used. The best fitting training data set was obtained with three neurons in the hidden layer, whih made it possible to predit gray values with auray at least as good as that of the experimental error, over the whole experimental range. On the validation data set, simulations and experimental data test were in good agreement (R>0.987). The developed model an be used for a reliable on-line state estimation and ontrol of roasting offee. Keywords: roasting offee, olor (grey level), neural networks 1. Introdution Coffee roasting is an unitary operation very important to develop the speifi organolepti properties (flavors, aromas and olour) whih, underlie the quality of offee and guarantee a good up of offee. Nevertheless, this proess is highly omplex, sine the quantity of heat transferred to the bean is ruial. During roasting offee, moisture loss and hemial reations (oxidation, redution, hydrolysis, polymerization, dearboxylation and many other hemial hanges), as well as major hanges (to olor, volume (swell), mass, form, bean pop, ph, density and volatile omponents) our, and CO 2 is generated. Finally, after these onsiderable hanges,
J. A. Hernándezr et al. the beans must be ooled rapidly to halt the reations (using water or air as a ooling agent) and prevent exessive roast whih will alter produt quality (Shwartzberg, 2000; Illy and Viani, 1998; Nagaraju et al., 1997; Raemy, 1981; Raemy and Lambelet, 1982; Singh et al., 1997; Sivetz and Desrosier, 1979). The quality of roasted offee is evaluated out-line using different parameters (for example: aroma, flavors, olor, bean temperature, ph, hemial omposition, bean pop, mass loss, gas omposition and volume) (Hernández-Pérez, 2002; Shwartzberg, 2000; Illy and Viani, 1998; Nagaraju et al., 1997). However, in the industrial setting, it is very diffiult to estimate these parameters on-line, and most ases the roast master has an essential role to play. He determines the operating onditions based on out-line measurements onerning organolepti properties (olor, aroma and flavors), and physial parameters (air temperature and stay time of the proess) (Hernandez et al., 2007). The proess is then adjusted for the nest bath. This method is only effetive if the quality of the raw material does not vary, whih is not the ase in the food industry. Thus in order to ontrol the proess, it must be possible to perform online measurements of produt quality. The grey level and surfae of the beans during the roasting offee are also two variables important to determiner the quality of roasting offee (Hernandez et al., 2007). However, these two variables are also diffiult to know on-line experimentally, for this raison it is neessary to estimate from equations mathematis. The progress of neurobiology has allowed researhers to build mathematial models of neurons to simulate neural behavior. Today s, neural networks are reognized as good tools for dynami modeling, and have been extensively studied sine the publiations of the pereptron identifiation method (Rumelhart and Zipner, 1985). Interest in these models inludes modeling without any assumptions about the nature of underlying mehanisms and their ability to take into aount non-linearities and interations between variables (Bishop, 1994). An outstanding feature of neural networks is their ability to learn the solution of the problem from a set of examples, and to provide a smooth and reasonable interpolation for new data. In the field of food proess engineering, reently, appliations of neural networks are arried out to orrelating olour to moisture ontent in the ooked beef (Qiao et al., 2007; Chaoxin et al., 2006). The aim of the work is to estimate the behaviour on-line of the grey level and the inrease in surfae of the beans from artifiial neural networks. In addition, this work will test the importane and effiieny of neural network to predit these two variables omplex. The model validation is made with an experimental database determined from image analysis system and the roasting offee proess. 2. Green offee Colombia green offee beans (Arabia) are roasted using a hot air flow as the heating medium. The experiments were arried out at a onstant air veloity of 4 m/s whih generated onstant air temperatures fixed by the roaster (190, 200, 210, 220, 230, 240,
J. A. Hernandez et al. 250, 260, 270, 280, 290 and 300 C) a period of 10 minutes. All experiments were performed in tripliate. 3. Green offee A stati roaster (SERVATHIN Series SV02 7817) was used to arry out the roasting experiments. A shemati draw of the roaster is given in Fig. 1. The offee bean were plaed on a mesh to keep them on stati suspension, where onvetion is the predominant mode of heat transfer. This offee roaster allowed us to equip it to ahieve the aim proposed. During roasting offee, the obtained diret measures were the air temperature in the roaster from a temperature aquisition system reported by Hernández et al., (2007) and the offee beans images by an image analysis system. 4. Data aquisition system Figure 1 shows the experimental system developed to follow the on-line olour [Red, Green and Blue] or the level of bright intensity (grey level) and the on-line surfae of the offee beans during roasting. This experimental system is onstituted as all the devie lassi of image analysis: A system of illumination: a soure of light with two small spotlight of fibre optis are establish for the illumination of the sene, A sensor of image: a amera video CCD (Charge-Coupled Devie) olour RGB (red, green and blue) SONY (XC-711P) working with an objetive 50 mm phi 25.5. This type of sensor is an of the less expense, A system of numeration: a omputer is used, whih is provided with one ard of aquisition video Hauppauge WinTV equipped with a onverter bttv 878, allowing the numeration of the image. We deliberately applied these equipments general publi for their robustness to lesser expense, but espeially the availability of the soure odes of the software pilots of the omponent bttv 878. and to the air temperature aquisition is given by the following: Thermoouples (type K) to measure the air temperature with a preision of +- 0.5 C, An Arom SCB7 thermoouple onditioner onneted to the personal omputer and, An Arom PCAD12/16H A/D onverter for the aquisition of air temperature in the roaster. The data aquisition and I/O port programming are written in C (Lawyer, 2000; Photis, 1999), on-line data proessing is done with a program written in Otave (Eaton, 2001) and an algorithm is managed in bash (Bourne, 2002).
J. A. Hernándezr et al. Figure 1. Aquisition sheme of experimental data used (olour and surfae). 5. Image Proessing System As shown the Figure 1, the amera video is installed outside the room of roaster and visualises the sene through a glassy window whih introdues a distortion negleted by the opti way. The images aquisition are made using the software bttvgrab (ommand bttvgrab -s1 -Q -l1 -oppm -dq -w640 -W480) (Walter, 2001). This software has the advantage to be used in a program written in bash. The images are saved onto the hard disk in format ppm (portable pixel map) at regular intervals of time (an image every twenty seonds) with a definition of 480 x 640 pixels in R,G,B, and grey level values. The images visualisation is arried out with the software xv. Therefore, the image system is onstituted by three stage: the aquisition (visualization of the objets, numeration), the treatment and the extration of information. It is important to notie that all these measures (air temperature, olour (RGB and grey level) and surfae) are aquired on-line allowing this way to take deisions of the proess in real-time. 6. Treatment and extration of the image information The main of the image treatment and extration is to have means to ontrol the online roasting proess onsidering parameters (olour and surfae), to determine the degree of roasting offee. In spite of the preautions whih we brought on the
J. A. Hernandez et al. measure, the images are taking in the onditions similar to the industrial way. However, it is diffiult of ontrol absolutely the illumination. Therefore, the images depend of the experimental onditions, beause there is a heterogeneity in the sene (every point of the image not reeive neither the same quantity nor the same quality of lighting). Consequently, in eah image obtained is onsidered their heterogeneity of the sene (Hernández Pérez, 2002). The information result is obtained by 3 matries (R,G,B). In order to working with this information (3x640x480) and to redue the time of alulation in omputer, the equation 1 is onsidered for obtained 1 matrix of 640x480: Red + Green + Blue grey = (1) 3 7. Color and grey kinetis experimental data Figure 2 shows the olour values (system [Red, Green, and Blue]) measured by the amera CCD and orreted during the roasting offee with air temperature fixed at 240 C. The urves behavior of the figure 2 are similar. All these urves present a behavior with a quik redution from 20 seonds, followed by an almost symmetrial growth (around 60-100 s), then, a redution uninterrupted of exponential behavior is determined. It is important to notie that the grey level is a variable important to determine the degree of roasting offee (Hernandez, et al., 2007b). Figure 2. Kineti of olour in Red, Green and Blue versus time for an air temperature fixed by the roaster of 240 C. For a better understanding of these urves behavior (fig. 2), (Hernández et al., 2007b) reported these urves behavior and the bean temperature funtion of time. The authors (Hernandez, et al., 2007b) desribed four different stages in the ourse of roasting offee proess: 1. During the first seonds (period from 0 to 20 s), beans remain the same olour, 2. When the bean internal temperature attained 100 C, the olor beomes lightly more dark (period about from 20 to 60 s),what an be owed to the vaporizing of not linked water,
J. A. Hernándezr et al. 3. Above 160 C, beans begin learing in a very important way (period about 60-100 s), 4. Then olour darkens little by little until that some offee harred (visually). Figure 3 represents the evolution of the grey level of the bean during roasting offee with different air temperatures. In high air temperatures (>260 C) the hange of the grey level is quiker during proess. The urves of grey level show therefore that the air temperature and time are two important fators for the proess of roasting. The repetition of tries, noted on the urves of grey level, shows a harater of allowable repetition well. 8. Experimental bean surfae kinetis Pixel size determination Distane on pitures is ounted in number of pixels. They depend therefore on experimental onditions and partiularly on the foal length of objetive and distane of objets in the amera. It is therefore neessary to play a alibration to onvert distane expressed in number of pixels, in millimetres. For this alibration, we arried out an image representing a irle of diameter 30 mm on a white bottom (Hernandez et al., 2007b). From irle determine easily the length and the height of the irle ommanding lines or olumns of the matrix of any element olorimetri of the image [R, V, B] or grey levels, as it is reported by Hernandez et al., 2007b. It an also determine the surfae of the limited square in the irle in number of pixels, to alibrate the surfae of this square. For it, we measured a length of 274 pixels and a height of 262 pixels for the image of the irle. Therefore, the square irumsribes of 900 mm 2 ount therefore 71788 pixels, onsequently, it dedut the surfae of a pixel in our system of 0.012537 mm 2. The surfae of the pixel is stoked in a text file whih will be later read in the ourse of the treatment of the images. Aquired results are similar whatever is the position of the irle on image. It is neessary to note that the length and the height should be equal, however it is not ase. In effet, this baby distortion an be owed to the spae resolution of the sensor CCD whih takes a sample more in width than in height.
J. A. Hernandez et al. Figure 3. Grey level kinetis for all experiments and their repetitions After the alibration of the surfae of the pixel, the bean surfae is measured in mm 2 from the number of pixels ounted on image. The figure 4 shows the inrease of the surfae of beans aording to the air temperature fixed by the roaster and time of
J. A. Hernándezr et al. roasting studied. These surfae urves (fig. 4) introdue an inrease of 15% at 70% for air temperatures fixed by the roaster between 190 C and 300 C. Moreover, Shwartzberg (2000) showed an expansion of volume of the bean of 50% at 120% for air temperatures of 270 C at 550 C with different roasted used. Dutra et al., (2001) determined an inrease of volume of 120% with time of roasting of 12 min at an air temperature of 275 C, using a diret heating. The air temperature is therefore one of the parameters key for the inrease of offee surfae in the ourse of roasting. It an also note that for air temperatures between 280 C at 300 C and for the upper time in 360 seonds, inrease is almost ended (see fig. 4), this an be owed to the optimum temperature of the bean whih is exeeded and onsequently the roasting is finished. Coste (1968) mentioned that the volume of the bean does not augment any more when bean temperature exeeds 280 C. These experimental data show that the offee bean begins to swell only when the bean internal temperature beomes the upper at 100 C (Hernéndez et al., 2007b). Figure 4. Experimental kinetis of inrease of the surfae for different air temperatures (190 C to 300 C). 9. Artifiial neural network Artifiial neural networks were inspired by the study of neurosienes. At present, they have appliations in the food industry (Qiao et al., 2007; Chaoxin et al., 2006) and notably in the speiality of the image analysis to define the quality of the produt
J. A. Hernandez et al. in real-time (Boillereaux et al., 2007; Park and Chen, 2000). Neural networks are able of learning the dynamis of proess from experimental data, onsidering nonlinearities of the system and orrelation between variables. As the natural neurons, they are determined to a great extent by onnetion between two elements, every onnetion between two neurons has a oeffiient (weight). This weighty notion allows to modulate the sign transmitted between two neurons aording to the state of link omputer synaptique whih links them up (Hernández Pérez, 2002). Neurons are put together in several layers interonneted to a given arhiteture. We used networks of lassial neurons of type pereptron multi-layer formed by three elements, typial for the approximation of funtions. These elements are formed by the input layer, hidden layer and output layer. Eah element of the layer is onneted to eah neuron input through the weight matrix. The best arhiteture is habitually determined by tries and errors. In order to alulate the stimulation S j of a neuron in the hidden layer, it is neessary to onsider the ativations A i of eah neuron of the input layer, whih is multiplied by their orresponding weight P ij. The bias B j is then added to regulate the threshold of ativation of the neurone (Demuth and Beale, 1998). j ( Pij Ai ) Bj S = + (2) The funtion of ativation is then applied to alulate A j, 2 A = j 1 1+ exp( 2 Sj) (3) These alulations are more simpler under matrix form (Dornier et al., 1998) S = P A + B (4) e ( S ) A = f (5) S = P A + B (6) s s s s ( S ) s s A = f (7) where A e, input standardized by the model; S, stimulation of the hidden layer, f, funtion of ativation (hyperboli tangent sigmoid transfer-funtion) (eq. 3), A, ativation of the hidden layer, S s, stimulation of the output layer, f s, funtion of ativation (linear transfer-funtion), to A s =f s. The oeffiients of network (weight P and bias B), the number of neurons in the hidden layer and the number of iterations of the algorithm of optimization are alulated in the training stage, minimizing a root mean square error of modelling in omparison with experimental data. The optimum model is the one who introdues minimal error. In this work, we used the toolbox for networks of neurons of software Matlab (Demuth and Beale, 1998) using an algorithm of optimization of type
J. A. Hernándezr et al. Levenberg-Marquardt, onsidered by Hagan and Menhaj (1994) as the being most effiient. To test the pertinene of our model, experimental database were split into learning and test database to obtain a good representation of the situation diversity. Two thirds in a learning database, whih will allow to alulate weights and biases optimum and a third in a test database whih will allow to validate the model testing its apaities of general implementation. The error on the learning database diminishes when the number of iterations augments. It also diminishes when the network augment the number of neurons in the hidden layer. But when network learns exatly the experimental learning database, the model it loses its apaities of general implementation on the test database. This phenomenon is alled over-fitting. The omparison between the root mean square error of the learning database and the root mean square error of the test database is a key riterion to optimize the number of iterations and avoid the over-fitting. 10. Grey level and surfae preditions The experimental data (grey level and surfae kinetis) were arried out at 12 different air temperatures fixed by the roaster (190, 200,... 300 C) with 3 repetions, the results 36 experimental for the grey level and 36 experimental for the bean surfae. From these database, we tried first of all to use kineti models reported in literature (Broyart et al., 1998; Krokida et al., 2001) to predit the grey level kinetis. However, we noted that these kineti models do not model experimental kinetis orretly. If we observed the grey level urves versus the time (fig. 3), it notie that the grey level kineti follows a tendeny ompliated with several stages. Rather than to onstrut a model ombining different laws, we propose an approah by neural networks. This work propose two neural networks models, the first alulating the grey level kinetis and the seond for surfae kinetis. To avoid taking into aount the variability of the produt of departure (different initial grey level), models apply to variables standardized as follows: redued _ grey(t) ng ng ng t t= 0 = ; t = 0 redued _ surfae(t) s s s t t = 0 = (8) t = 0 where ng t et ng t=0 are the grey level in time and initial value, respetively, similarly for the surfae s t and s t=0. Aording to previous results (Hernández Pérez, 2002), we planned to use the bean temperature and time as two input variables for the first model (redued grey-level) and the variables of air temperature and time as two input variables for the seond model (projeted surfae). The best results for the grey level are obtained from two input variables: bean temperature simulated by the dynami model T b proposed by Hernández et al., (2007a), and the time of roasting t (see fig. 5a).
J. A. Hernandez et al. It an also plan to use the bean temperature experimental, but this solution is not realisti at the industrial level beause it is very diffiult to obtain in a roaster. For the seond model (redued surfae), it was onsidered the air temperature fixed by the roaster T a and the roasting time t (fig. 5b). In order to identify the oeffiients of the two models and to validates its, we divide the experimental results in a base of learning made up of eight experiments (whih are 190, 210, 220, 240, 250, 270, 280 and 300 C) and a base test omposed of four kinetis (200, 230, 260 and 290 C). Eah experiment ontains three repetitions. Figure 5. Neural networks, with the roasting time as variable of input. 11. Results and disussion of the two models Two models of artifiial neural network are used to predit the variations of grey level starting from the bean temperature simulated and expliit roasting time. The another predited inrease in the surfae of the grain, aording to the air temperature fixed by the roaster and expliit roasting time. Grey level model In the phase of learning, the best network to predit the grey livel omprises 3 neurons in hidden layer. It thus has 13 oeffiients (9 weights and 4 bias). To validate this model, we simulated the grey level kinetis ontained in the test database. The evolution of the redued grey level experimental for an air temperature fixed by the roaster of 260 C is ompared with the grey simulated values (see fig. 6b). This model predits in a satisfatory way all the urves of the grey level during roasting with a
J. A. Hernándezr et al. oeffiient of orrelation R=0.987. It is important to note that the model predited well the first phases of the urve (between 0 and 100 seonds) whih is omplex. Moreover, the layout of all the values simulated aording to the experimental data of the test database (fig. 6) shows a balaned distribution of the residues. The standard deviation for the test database is of 0,0229 and for the training of 0,0256. The similarity of these two values shows the preditive apaity of this model. Figure 6. Testing of the neural model for the grey level. (a) redued grey level funtion time at air temperature of 260 C: (*) experimental data (-) simulated data with three repetitions. (b) the same redued grey level onverted in grey level: experimental data (*) and simulated data (-). () experimental data funtion of the values simulated for all the test database. Surfae model In order to predit the bean surfae kinetis, the best network is similar with the preedent. The Figures 7a and 7b present the evolution of inrease in surfae experimental and simulated aording to time for kinetis ontained in the test database (260 C). It an note the variation of the bean surfae during roasting (fig. 7b). The preision of the model is onsidered to be satisfatory, beause the oeffiient of orrelation is of R=0.993 if the experiments of the two bases are onsidered. As for the grey level, the figure 7 ompare the values simulated with the experimental values for all the test database, it shows the apaity to predit the urves of inrease in not learned surfae. This is onfirmed by the omparable standard deviation on the test database (0.022) and of learning (0.030).
J. A. Hernandez et al. Figure 7. Testing of the neural model for the inrease of bean surfae during roasting. (a) redued surfae funtion of the time at air temperature of 260 C: (*) experimental data (with two repetitions) and (-) simulated data. (b) the same redued surfae onverted in inreased of surfae (%): experimental data (*) and simulated data (-). () experimental data funtion of the values simulated for all the test database. 12. Conlusion This study proposes two artifiial neural networks models, whih predit the grey level and the surfae of the bean during roasting offee. The two neural networks models were suessfully trained with experimental database and validated with a fresh database (in the speified range of key operating onditions), obtained a R >99 %. These models onsider the simulated bean temperature, air temperature fixed by the roaster and roasting time as variables of input. It is important to note that the grey level and surfae are two parameters very important to determiner the quality on-line. Finally it is possible to obtain the quality of the roasting offee from these parameters: grey level and surfae from the two proposed neural networks. In addition, it is important to notie that the dynami model, whih predit the bean temperature reported by Hernandez et al., (2007a) is onsidered as a input variable of the neural network model to predit the grey level kinetis. Referenes Bishop, C. M. (1994). Neural networks and their appliations. Review Sine Instrument, 65, 1803-1832. Bourne, S. (2002). Bash (free software). In Bourne Again SHell. [On-line] Available from internet http://www.gnu.org/software/bash/bash.html, onsulted April 2005. Free Software Foundation. Boillereaux, L., Cadet, C. & Le Bail, A. (2007). Thermal properties estimation during thawing via real-time neural network learning. Journal of Food Engineering, 57, (1), 17-23.
J. A. Hernándezr et al. Broyart, B., Trystram, G. & Duquenoy, A. (1998). Prediting olour kinetis during raker baking, Journal of Food Engineering, 35, 351-368. Chaoxin, Z., Da-Wen, S., & Liyun Z. (2006). Correlating olour to moisture ontent of large ooked beef joints by omputer vision. Journal of Food Engineering, 77, (4), 858-863. Coste, R. (1968). Le aféier. G. P. Maisonneuve et Larose, XIV (ed), 11, rue Vitor- Cousin, Paris (V e ). Demuth, H. & Beale, M. (1998). Neural Network Toolbox for Matlab, user's guide version 3. The MathWorks In, USA. Dornier, M., Heyd, B. & Danzart, M. (1998). Evaluation of the simplex method for training simple multilayer neural netwoks. Neural omputing and appliations, 7, 107-114. Dutra, E. R., Oliveira, L. S., Frana, A. S., Ferrez, V. P., & Afonso, R. J. (2001). A preliminary study on the feasibility of using the omposition of offee roasting exhaust gas for the determination of the degree of roast. Journal of Food Engineering. 47, 241-246. Eaton, J. W. (2001). Otave (free software). In GNU Otave. [On-line] Available from internet http://www.otave.org/otave.html, onsulted April 2007. University of Wisonsin. Guyot, G. (1993). Torréfation: meanismes, transformation physiques et himiques. In Journées du CAFE,CIRAD-CP á Montpellier. Frane. Hagan, M. & Menhaj, M. (1994). Training feedforward networks with the marquardt algorithm. IEEE transations on neural networks, 5,(6), 989-993. Hernández, J. A., Heyd, B., Irles, C., Valdovinos, B., & Trystram G. (2007a). Analysis of the heat and mass transfer during offee bath roasting. Journal of Food Engineering. 78, 1141-1148. Hernández, J. A., Heyd, B & Trystram, G. (2007b). On-line quality estimation during offee roasting: Part I~-~olor (gray) and surfae kinetis. Submitted to Journal of Food Engneering. Hernández Pérez, José Alfredo (2002). Étude de la torréfation: modélisation et détermination du degré de torréfation du afé en temps réel. Ph. D. Thesis in Éole Nationale Supérieure des Industries Agrioles et Alimentaires. Frane. Illy, A., & Viani, R. (1998). Espresso Coffee. Aademi press, San Diego, CA 92101, U.S.
J. A. Hernandez et al. Krokida, M. K., Oreopoulou, Z. B., & Marinos, D. (2001). Colour hanges during deep fat frying, \textit{journal of Food Engineering} 48, 219-225. Lawyer, S. D. (2000). Serial-HOWTO. In the Linux Serial HOWTO. [On-line] Available from internet http://www.linuxdo.org/howto/serial-howto.html. Consulted January 2005. Nagaraju, V. D., Murthy, C. T., Ramalakshmi, K., & Srinivasa, R. P. N. (1997). Studies on roasting of offee beans in a spouted bed. Journal of Food Engineering, 31, 263-270. Park, B., & Chen, Y. (2000). Real-time dual-wavelength image proessing for poultry safety inspetion. Journal of Food Proess Engineering, 23, 329-351. Photis, M. G. (1999). Coffee-HOWTO.txt. In software offee-howto, [On-line] Available from internet http://www.faqs.org/dos/linux-mini/coffee.html\# ss1.1, onsulted January 2007. University Cambrige. Qiao, J., Ngadi, M. o., Wang, N., Gariepy, C. & Prasher, S. O. (2007). Pork quality and marbling level assessment using a hyperspetral imaging system. Journal of Food Engineering, 83, (1), 10-16. Raemy, A. (1981). Differential thermal analysis and heat flow alorimetry of offee and hiory produts. Thermohimia Ata, 43, 229-236. Raemy, A., & Lambelet, P. (1982). A alorimetri study of self-heating in offee and hiory. Journal of Food Tehnology, 17, 451-460. Rumelhart, D. & Zipner, D. (1985). Feature disovering by ompetitive learning. Cognitive Siene, 9, 75-112. Singh, P., Singh, R., Bhamidipati, S., Singh, S., & Barone, P. (1997). Thermophysial properties of fresh and roasted offee powders. Journal of Food Proess Engineering, 20, 31-50. Sivetz, M., \& Desrosier, N. W. (1979). Coffee tehnology. In AVI Publishing Co (pp. 226-264). Westport. Shwartzberg, H. G. (2000). Modelling bean heating during bath roasting of offee beans. In \textit{engineering and Food for the 21 st Century, edited by J. Welti- Chanes, G. Barbosa-Canovas, JM Aguilera, CRC Press LLC, London, New York, Boa Raton. Walter, J. (2001). bttvgrab-0.15.10.tar.gz (free software). In bttvgrab, [On-line] Available from internet http://www.garni.h/bttvgrab/, onsulted January 2007.