Assessment of fresh pork color with color machine vision 1

Transcription

1 Assessment of fresh pork color with color machine vision 1 F. J. Tan,2, M. T. Morgan*, L. I. Ludas,3, J. C. Forrest,4, and D. E. Gerrard Departments of *Agricultural and Biological Engineering and Animal Sciences, Purdue University, West Lafayette, IN ABSTRACT: Currently, fresh pork color is visually evaluated using either the Japanese Pork Color Standards (JPCS) or the National Pork Producers Council Pork Quality Standards (NPPC) as a reference. Although useful, visual evaluation of meat color can vary with evaluator and may be quite expensive. In this study, three separate studies were used to compare the ability of color machine vision (CMV) and untrained panelists to evaluate pork color. Panels visually evaluated over 200 pork loin chops using either the JPCS or NPPC reference standards. Results from each panel were used to evaluate the ability of the CMV to sort pork loin chops based on the same criteria. Representative samples, typical of each color class, were used to train neural-network-based image processing software. After training, the CMV system was used to evaluate quality classes of pork samples based on color distribution. Classification by CMV was compared with the average panel score, rounded to the nearest integer. Training the CMV system using images of actual meat samples resulted in a stronger correlation to panel scores than training with either set of artificial color standards. Agreement between the CMV system and the panels was as high as 90%. Agreement between individual panelists and the integer panel average (52 to 85%) was less than that observed for CMV classification. Finally, the on-line performance of CMV using a laboratory conveyor system was simulated by repeatedly classifying 37 samples at a speed of 1 sample per second. Collectively, these results demonstrate that CMV is a rapid and repeatable means of evaluating pork color. Key Words: Color, Quality, Pork 2000 American Society of Animal Science. All rights reserved. J. Anim. Sci : Introduction Demand for high-quality pork by domestic and export markets has created a heightened awareness of product quality for U.S. pork producers. The Pork Chain Quality Audit (Meeker and Sonka, 1994) indicated that variations in pork quality and composition account for over 50% of lost opportunities to the industry. A survey by Kauffman et al. (1992) found that only 16% of the carcasses have ideal lean quality based on color, firmness, and water-holding capacity. If the greatest opportunities for increased sales lie in the export market where high-quality pork is demanded, efforts must be made to improve pork quality. 1 Purdue Univ. Agric. Res. Programs Journal Paper No. 16,112. This study was supported in part by the National Pork Board in cooperation with the National Pork Producers Council. 2 Present address: Dept. of Anim. Sci., The Ohio State Univ., Columbus, OH Present address: 1141 Budapest, Fogarasi ut 135, Hungary. 4 Correspondence: 1151 Smith Hall, ( jforrest@ purdue.edu). Received October 18, Accepted August 18, Color may be the most important factor that influences the appearance and attractiveness of pork to consumers (Faustman and Cassens, 1990). Unfortunately, the appearance of pork at the retail level varies across and within packages. Although sorting could reduce some of this variation, this underscores the problem of product uniformity in the industry. Unfortunately, a human-based evaluation and classification system relies heavily on visual perception of color that may not be sufficiently accurate for the industry (Daley and Thompson, 1992). Development of an objective means for assessing pork quality would allow quality to enter a value-based pricing system. In addition, such technology would allow for greater segregation of quality at both the packer and retail levels and provide a means of feedback to producers. Automated systems are needed to achieve accurate, consistent, objective evaluation of meat quality with minimal labor and operate at line speeds to sort primal cuts. Recent advancements in color machine vision (CMV) systems provide an approach to achieve these goals (Tillett, 1991; Precetti, 1994; Gerrard et al., 1996). Therefore, the objective of this study was to determine the ability of a CMV system to predict visually assessed color scores of fresh pork. 3078

2 Pork color and machine vision 3079 Table 1. Number of panelists and loin samples and lighting conditions for each color study Study Location No. of panelists No. of samples Lighting conditions 1 Purdue University 6 73 Fluorescent 2 Iowa State University Tungsten halogen 3 University of Illinois Tungsten halogen Materials and Methods The ability of the CMV system to predict visually assessed panel scores was evaluated in three separate studies by comparing computer color classification of loin chops with visually assessed color scores assigned by untrained panels. In the first study, a panel used the five color classes defined by the National Pork Producers Council (NPPC; NPPC, 1991) as a means to score pork color. In the second and third studies, the six color classes based on the Japanese Pork Color Standards (JPCS) (Nakai et al., 1975) were used. In the second study, the panel re-scored samples so that repeatability could be evaluated, whereas in the third study both the CMV system and panelists re-scored the same loin chops. Finally, accuracy and repeatability of the CMV system was evaluated during real-time classification of pork loin chops moving on a conveyor belt. In this case, accuracy refers to agreement between CMV classification and average panel score rounded to the nearest integer. Pork Samples In each of the three studies, 73, 79, and 51 pork loin chops, respectively, were evaluated for color by the CMV system and an untrained panel. Loin chops were fabricated from different loins adjacent to the 10th and 11th ribs of carcasses obtained from the Purdue University Meat Laboratory (Study 1), the Iowa State University Meat Laboratory (Study 2), and the University of Illinois, Champaign-Urbana Meat Laboratory (Study 3). For the first two studies, bone-in, untrimmed chops were evaluated in the cooler after a 24-h carcass chill at 5 C and a 10-min bloom time following fabrication. In contrast, boneless, 4-mm trimmed chops were used for the third study. Some chops in Study 3 were collected from carcasses treated with epinephrine 12 h prior to slaughter (McCaw et al., 1997) or held at 30 C after slaughter (McCaw et al., 1994) to obtain more variation in pork color. s and Scoring Meat scientists with extensive experience in evaluating meat quality characteristics, particularly fresh pork color, and representing both academia and industry served as untrained panelists. Panels visually assessed pork loin chops and assigned color scores. The Japanese (Nakai et al., 1975) or NPPC (NPPC, 1991) color standards were available to panelists in all three studies. Pork color is typically assessed and sorted by in-house evaluators who may or may not have undergone extensive training; if trained, the training is not uniform from company to company. Therefore, this untrained expert panel more nearly represents the variation and expected lack of reproducibility that may be encountered across the industry. Bloomed chops were presented to the panelists unwrapped in trays. In the first study, a color score for each sample was marked on a color-class ruler on which pale pinkish gray, grayish pink, reddish pink, purplish red, and dark purplish red were assigned to color scores 1 to 5, respectively (NPPC, 1991). A score of 3.0, for example, represented the lightest color of class 3. If the panelist marked on the line between two color scores, interpolation was used to determine the score to the nearest tenth. Scores for each of six panelists were then averaged and rounded to the nearest integer. These integer values were used as the final color score for each sample. In the second and third studies, integer color scores (1 to 6) were assigned to each sample by the panelists according to the JPCS standards (Nakai et al., 1975). Figure 1. Color machine vision system.

3 3080 Tan et al. Image Acquisition Figure 2. Application procedures for using SAMPLEX. The Japanese color standards were considered to be the lightest color that would still be associated with each class. Final color scores for each sample were again assigned by rounding the average score of the panelists to the nearest integer. To evaluate the repeatability of the visual panel, samples were re-scored by all panelists, 15 to 20 min later, in random order. Selected samples scored similarly by panelists in each round were used as a training set to teach the neural network software of a CMV system. Study 1 used eight standard 40-W (2,910 lumens, 4,100 K) fluorescent lights. Lighting in study 2 consisted of four 90-W halogen flood lights (1,280 lumens, 2,900 K). In the third study, three 500-W tungsten halogen (11,100 lumens, 3,000 K) lights were used so that panelists evaluated samples under similar conditions as the CMV system. Table 1 summarizes the number of samples, number of panelists, and lighting conditions in each of the three studies. A schematic of the CMV system used in this study is shown in Figure 1. An RGB camera (Sony, New York, model XC-711 CCD) with 20-mm focal length was used as the vision sensor. The camera was positioned 44 cm directly above the samples. The RGB analog signals were captured, processed, and stored by image processing hardware (Matrox Electronic Systems Ltd., Quebec, Canada, models IM-640, IM-CLD, and IM- RTP) installed in a personal computer. Lighting consisted of two external, 500-W tungsten halogen bulbs (GE Quartzline, Q500T3/CL, 3,000 K) directed at a 45 angle to the surface of the samples. A light box was built to surround the camera s viewing area and eliminate the effects of ambient light. Pork samples were placed on a non-glare black plastic plate in the field of view. Background and fat colors not relevant to lean color were eliminated with color classification. For real-time classification, the light box was mounted over a conveyor belt. Belt speed was fixed at 0.3 m/s. The camera was set to capture an image every time a sample of meat (manually positioned on the end of the conveyor on a styrofoam tray) passed by an optical sensor. Camera shutter speed was set at 1/125 s and the aperture at f 8. In all evaluations, images were automatically classified using a neural network classifier in software developed at Purdue University. Color Classification Software (SAMPLEX) For objects that are very uniform in color, or have little variation in color distribution, simple threshold methods are used to assign a color classification (Barni et al., 1997). However, these methods are not as useful for objects, such as meat, with a more complex color distribution. For example, a pork chop scored as a number 2 (according to the NPPC color scale) may contain colors from other classes (such as color class 1, 3, and even 4) dispersed over the surface of the chop. Furthermore, cuts with the same color score may possess different distributions of each color across the surface. In order to classify these complex color distributions more accurately, an advanced two-step, neural networkbased color classification software was created. This Table 2. Typical agreement matrix comparing samples scored in each class by panelist C and the entire panel (Study 1) Entire panel a Class 1 b Class 2 Class 3 Class 4 Class 5 Class Class C Class Class 4 4 Class 5 0 a Mean panel scores rounded to the nearest integer. b Class 1 represents pale pinkish gray ; class 5 represents dark purplish red.

4 Pork color and machine vision 3081 Table 3. Percentage agreement between individual panelist and panel for each quality class (Study 1) Class 1 Class 2 Class 3 Class 4 A 88 (7/8) a 68 (32/47) 75 (6/8) 100 (4/4) B 29 (2/7) 74 (35/47) 88 (7/8) 100 (4/4) C 88 (7/8) 64 (34/53) 100 (8/8) 100 (4/4) D 63 (5/8) 70 (35/50) 83 (5/6) 100 (4/4) E 75 (6/8) 87 (46/53) 29 (2/7) 100 (4/4) F 63 (5/8) 79 (41/52) 44 (4/9) 100 (4/4) a Number of samples classified by panelist that agreed with the panel group/number of samples scored by panelist. software, SAMPLEX (Purdue Research Foundation, 1995) contains a color image display, classification algorithms, and a classified image display. Figure 2 illustrates the functioning of the SAMPLEX software. A brief description of the software is given here and more details are presented in Precetti (1994). First, in order to create a neural network classifier that performs similarly to visual evaluation, a training set is used that includes samples that have been visually scored to some reference standards. The first-step classifier of SAMPLEX is used to recognize the different colors (i.e., different hue, saturation, and value combinations) that are normally found in a pork chop. During training, standard colors are chosen from either the reference standard images (NPPC, 1991) or meat sample images in which areas can be selected to represent each color group. Inside each sampled area, one color class is assigned to all picture elements (pixels). The red, green, and blue values of each pixel within this sampled area are recorded in a database along with the color class assignment. This database containing all pixel information is then used to train the step 1 neural network. The first step of the classifier segregates each image pixel into either a lean color class (JPCS = 1 to 6; NPPC color standards = 1 to 5), a fat class, or background class. This first-step classification results in the percentage of pixels in each class. The second-step classifier, or quality classification step, then evaluates the percentage color distribution over the entire image. Using results of the first-step classifier and the panel scores for each sample in the training set, the second-step classifier is trained to ana- lyze a sample image and assign it to a color quality class. From this point forward, a color quality class will be referred to as a quality class. Therefore, the end result of the second-step classification is a quality class number for each image. After training the software based on either the NPPC color standards or the JPCS, the two-step color classifier was used to sort pork loin chops. Performance Evaluation As explained earlier, the final color score for each sample was the mean panel score, rounded to the nearest integer. Computer-classified scores or quality classes were then compared to these panel scores. An agreement matrix was used to evaluate the performance of each panelist and computer classifier. The number in cell ij (i th row and j th column) of the agreement matrix is the number of samples given the i th color score by the panelist, or CMV classifier, assigned to the j th class by the mean of the entire panel. The sum of diagonal numbers divided by the total number of samples is defined as the percentage agreement between the column and row methods. Results and Discussion Study 1 (Based on NPPC Standards) Comparison between Untrained Individual Scores and the Entire Panel. Table 2 illustrates one agreement matrix that shows 73% of samples (53/73) classified by panelist C were given the same score as Table 4. Agreement between individual panelist and panel for each round (Study 2) Round A B C D E F a G H I J K L M 1 No. of samples Agreement, % No. of samples Agreement, % a F did not score round 2.

5 3082 Tan et al. Table 5. Repeatability of individual panelist and panel (Study 2) A B C D E G H I J K L M No. of samples: Panel a Repeatability, % b a Based on mean of entire group. b Repeatability of rounded average color scores between scoring rounds. the average score assigned by the entire panel. Table 2 shows that there were 8 class 1, 53 class 2, 8 class 3, and 4 class 4 loin chops in the study as determined by the panel average scores. The percentage agreement between individual panelists and average integer of the group ranged between 72 and 80% (mean of 74% ± 2.5). Although percentage agreement for all panelists was high, individual panelists performed differently when evaluating different quality classes. Table 3 shows the percentage agreement between individual panelists and the entire panel for each of the quality classes. The number of samples scored similarly by individual panelists and the whole panel was divided by the total number of samples in each quality class. This demonstrated that some panelists had difficulty, or bias, identifying certain quality classes. For example, panelist B had difficulty recognizing samples belonging to the NPPC quality class 1 (pale pinkish gray). Scores by panelist B matched the panel average for only two out of seven class 1 samples. Likewise, even though the total percentage agreement for panelist E was the highest, scores for only two out of seven class 3 samples matched the panel score. This corresponds to 29% agreement for quality class 3 and shows that panelist E had difficulty recognizing quality class 3 samples. This variation asso- Figure 3. Agreement percentages between computer classifier and panel group scores (Study 2). Classifier numbers refer to the number of images used during training (i.e., classifier 753 was trained with seven class 2 images, five class 3 images, and three class 4 images during step 2 training). ciated with visual color scoring is a limitation when training a CMV system. Computer Classifiers using SAMPLEX. The first classifier (classifier 1) created for this study used images of the NPPC Standards (NPPC, 1991) to train the twostep algorithm in SAMPLEX. Standards were sampled to train the step 1 neural network and then the entire image of each standard (NPPC color scores 1 through 5) was used to train the step 2 neural network. This first classifier was then used to classify 73 meat samples. Computer quality scores were then compared to mean panel scores. The agreement between this first classifier (classifier 1) and the panel scores was only 58%. This demonstrated that training the CMV system with printed image standards does not result in a satisfactory classifier. Six additional classifiers were then created using an increasing number of images of real meat samples during training. The best of these classifiers used images of real meat for defining each of the colors (1 to 4) and the image of standard 5 for the fifth color in step 1 training (a sample representative of quality class 5 was not available). Then, for step 2 training, four images of real meat were used to represent each of the four quality classes (1 to 4) and the NPPC standard was used for quality class 5. Agreement between the best of these classifiers and panel scores was 81%. These results demonstrated that training a classifier with images of real meat resulted in a CMV system with greater agreement with the panel than each of the individual untrained panelists. Upon inspection of the results, however, there were still some obviously misclassified areas on several of the meat images, mostly in areas containing bone, fat, shadows, and hide. As a result, images were subsequently trimmed using an image processing package and reclassified using the same classifier. Results of classifying modified images improved agreement to 85%. This improvement suggests that image segmentation prior to color classification would improve CMV system accuracy. Automatic segmentation of muscles in an image is possible but adaptation to pork loin chops would be necessary (Gerrard et al., 1996). Study 2 (Based on Japanese Pork Color Standards) In this study, there were 21 samples in quality class 2, 52 samples in quality class 3, and 6 samples in quality class 4. Unfortunately, no meat samples were placed

6 Pork color and machine vision 3083 Table 6. Percentage agreement of individual panelist and computer classifier with the visual panel group without the 23 images used for training the classifier (Study 2) Computer A B C D E F G H I J K L M classifier a Agreement, % a Neural network classifier trained in step 2 with 10 class 2, 10 class 3, and 3 class 4 images of real meat. in classes 1, 5, or 6, using the average integer score of the panelists. Table 4 shows the agreement between each of the 13 panelists in this study with the group. In this study, the panelist s agreement with the group scores ranged from 30 to 89%. Repeatability of Individual s. To evaluate the repeatability of the panel, all samples were re-scored by 12 of the 13 panelists. These results were then compared to determine the repeatability of the panel scores as a group. Table 5 illustrates that the repeatability of individual panelists ranged from 63 to 81%. For example, only 63% of the samples received the same score from panelist B in both rounds of scoring. However, as a group the repeatability was 92% between rounds. Only 6 of the 79 samples were re-scored differently in each round based on the average panel scores. Evaluation of Computer Classifier Performance. In this study, 13 computer classifiers were developed that contained the same step 1 neural network but not the same step 2 neural network. Sampling portions of real meat images that were classified into quality classes 2, 3, or 4 by the panel created the step 1 neural network. Real meat was used because previous attempts to train a neural network using the JPCS color chips yielded poor results. The first classifier, 111, used one image for class 2, one image for class 3, and one image for class 4 for training in step 2. Although all three of these images agreed with the panel results, they may not have represented all the possible distributions of color found in typical samples. Thirty-five out of 52 class 3 samples (67%) were scored as class 2 using classifier 111. This implies that classifier 111 had difficulty differentiating meat samples between color classes 2 and 3. This mis-scoring by classifier 111 yielded a total agreement of only 49% with the panel. After adding three more images to the training set (one image per color class) to create classifier 222 (two images used for training each class 2, 3, and 4), the agreement between the computer classifier and the panel score only increased to 51%. After adding another class 2 sample, creating classifier 322, the accuracy improved to 79%. Figure 3 illustrates that adding even more images to the training sets (classifiers 332 through 1083) only slightly improved the accuracy of the computer classifiers. Classifier was trained with 10 images for classes 2 and 3 and three images for class 4, for a total of 23 images. This classification resulted in the highest accuracy (90%). Comparing the classification results of classifier with those of the panel showed that it performed better than any individual panelist. Even without considering the 23 images used during training (i.e., removing them from the evaluation set, leaving 56 samples), computer classifier still performed better (86%) than each individual panelist scoring those same samples (57 to 84%; Table 6). This demonstrated that the classifier performed well enough to sort samples that were not used during training of the neural network. Study 3 (Based on Japanese Pork Color Standards) This study was similar to study 2 described above, except that the CMV system was used to capture two images of each loin chop to also assess the CMV repeatability. In this study, there were 8 samples in class 1 in each visual assessment round, 3 and 7 class 2 samples, 21 and 17 class 3 samples, 12 and 14 class 4 samples, and 8 and 6 class 5 samples in rounds one and two, respectively. A total of 51 samples from 51 different loins were evaluated in each round. Table 7 shows the agreement between each panelist and the entire panel. Agreement levels ranged from 42% to 84%. Table 8 shows the repeatability of each panelist between round 1 and 2 scoring of the same loin chops. Because the repeatability of some of the panelists was so low, only scores from those panelists who had aboveaverage repeatability were used to compare with the CMV systems classification results. This screening left 10 panelists who had repeatabilities above 70%. The repeatability of this 10-member panel was increased to 96%. Because this study contained meat samples with scores ranging from 1 to 5 on the Japanese color scale, Table 7. Agreement between individual panelists and panel (Study 3) Agreement, %

7 3084 Tan et al. Table 8. Repeatability of individual panelist and group (Study 3) 1 2 b 3 b Panel a Repeatability, % c a Consisted of the mean score of 10 panelists with above 70% repeatability. b Based on 36 samples scored in each round. c Repeatability of integer panel averages between scoring rounds. a new SAMPLEX classifier was created by sampling colors from real meat images in the first step and using four images for each quality class to train step 2. Figure 4 shows the number of meat samples classified into each quality class in each round by the 10-member panel and the computer classifier. Repeatability of the computer classifier was 98% with only one disagreement between rounds. This demonstrated that the repeatability of the CMV system was much higher than the average panelist (80%) or even the most consistent panelist (90%). Table 9 shows the agreement matrix between the panel classification and the computer classifier. The agreement between panel and the computer classification in each round was only 77%. There were 39 agreements out of 51 samples in each round. This is low, but the average agreement between individuals and the panel was only 66%. Therefore, the CMV classifier was 10 percentage points better than the average panelist. The lower than expected agreement between the CMV system and the panel may indicate that more images need to be used for training. However, when examining the agreement matrix, only one class 5 sam- ple was misclassified by more than one color class in round 1. All other errors were only off by one color class. Real-Time Color Classifier Evaluation. The main objective for the real-time testing was to determine whether the hardware and software of the CMV system could operate while samples were moving on a conveyor belt. Although the system was not tested in a processing plant environment, samples were placed on the conveyor at a speed of approximately one sample per second to simulate on-line operation. The same classifier used in Study 3 was modified to add a shadow class to the existing five lean color classes plus background and fat classes in the step 1 neural network. Thirty-seven loin chops were classified a total of six times each, three stationary classifications and three real-time classifications. Stationary and real-time classifications were performed in alternating rounds with random sample order. Results showed that 29 out of the 37 samples received the same quality class score in each of the three stationary rounds. For six samples, one of the three scores differed by one quality class and the remaining two samples received one of three scores that differed by Figure 4. Number of samples scored into different groups (1 to 5) by the panelists and computer classifier in two rounds of evaluation. Within a color grade, bars represent, from left to right, scores from rounds 1 and 2 of the panel evaluation and scores from rounds 1 and 2 of the computer evaluation, respectively.

8 Pork color and machine vision 3085 Table 9. Agreement matrix of samples classified by the panel and the computer classifier in two rounds (Study 3) Panel Classifier Class 1 Class 2 Class 3 Class 4 Class 5 Class 1 7/7 a 1/1 0/0 0/0 0/0 Class 2 0/0 2/4 2/0 0/0 0/0 Class 3 0/0 0/2 18/17 4/4 1/0 Class 4 0/0 0/0 1/0 5/5 0/0 Class 5 0/0 0/0 0/0 3/5 7/6 a Number of samples scored in round 1/number of samples scored in round 2. two class scores. Therefore, of the 111 stationary classifications there were eight repeatability errors corresponding to 93% stationary repeatability. During the real-time evaluations, 35 out of 37 samples received the same quality class score in each of the three rounds, or 95% real-time repeatability. Both of the repeatability errors were misclassifications of only one quality class. Comparison of real-time and stationary classification results revealed that 4 out of the 37 samples were misclassified, corresponding to 89% agreement between the real-time and stationary classifications. Implications A color machine vision system that was trained using images from pork classified by a visual panel based on NPPC or JPCS standards was capable of classifying pork loin chops up to 86% agreement with visually assessed panel scores. The color machine vision system performed with more consistency and at a higher percentage agreement with the panel than most of the untrained individual panelists, demonstrating the utility of color machine vision for evaluating pork color. Used with an effective tracking system, this technology could potentially sort retail meat cuts into uniform quality/color groups before shipping to retail merchandisers. Furthermore, a CMV system can operate at simulated on-line speeds with accuracy and repeatability similar to stationary evaluations. Literature Cited Barni, M., V. Cappellini, and A. Mecocci Colour-based detection of defects on chicken meat. Imag. Vis. Comput. 15: Daley, W. D., and J. C. Thompson Color machine vision for meat inspection. In: Proc. ASAE Conf. Food Processing Automation II Conf., St. Joseph, MI. pp Faustman, C., and R. G. Cassens The biochemical basis for discoloration in fresh meat: A review. J. Muscle Foods 1: Gerrard, D. E., X. Gao, and J. Tan Determining beef marbling and color scores by image processing, J. Food Sci. 61: Kauffman, R. G., R. G. Cassens, A. Scherer, and D. L. Meeker Variation in pork quality: History, definition, extent, and resolution. National Pork Producers Council, Des Moines, IA. Meeker, D. L., and S. Sonka Pork Chain Quality Audit. National Pork Producers Council, Des Moines, IA. McCaw, J. C., M. Ellis, M. S. Brewer, and F. K. McKeith Incubation temperature effects on physical characteristics of normal, dark, firm and dry, and halothane-carrier pork longissimus. J. Anim. Sci. 75: McCaw, J. C., F. K. McKeith, T. R. Carr, and M. Ellis The effects of muscle condition, temperature, and ph on the color and WHC of fresh pork. In: Proc. 47th Recip. Meat Conf., University Park, PA. p 85 (Abstr.). Nakai, H., F. Saito, T. Ikeda, S. Ando, and A. Komatsu Standards models of pork color. Bull. Natl. Inst. Anim. Indust. 29: NPPC Procedures to Evaluate Market Hogs. 3rd ed. National Pork Producers Council, Des Moines, IA. Precetti, C. J Advanced classification system for biomaterials. Ph.D. dissertation. Purdue University, West Lafayette, IN. Purdue Research Foundation Copyright SAMPLEX software. West Lafayette, IN. Tillett, R. D Image analysis for agricultural processes: A review of potential opportunities. J. Agric. Eng. Res. 50: