Abstract

A hyperspectral-multispectral line-scan imaging system was developed for differentiation of wholesome and systemically diseased chickens. In-plant testing was conducted for chickens on a commercial evisceration line moving at a speed of 70 birds per minute. Hyperspectral image data was acquired for a calibration data set of 543 wholesome and 64 systemically diseased birds and for a testing data set of 381 wholesome and 100 systemically diseased birds. The calibration data set was used to develop the parameters of the imaging system for conducting multispectral inspection based on fuzzy logic detection algorithms using selected key wavelengths. Using a threshold of 0.4 for fuzzy output decision values, multispectral classification was able to achieve 90.6% accuracy for wholesome birds and 93.8% accuracy for systemically diseased birds in the calibration data set and 97.6% accuracy for wholesome birds and 96.0% accuracy for systemically diseased birds in the testing data set. By adjusting the classification threshold, 100% accuracy was achieved for systemically diseased birds with a decrease in accuracy for wholesome birds to 88.7%. This adjustment shows that the system can be feasibly adapted as needed for implementation for specific purposes, such as paw harvesting operations or prescreening for food safety inspection. This line-scan imaging system is ideal for directly implementing multispectral classification methods developed from hyperspectral image analysis.

INTRODUCTION

Since the 1957 passing of the Poultry Products Inspection Act, USDA inspectors have been inspecting all chickens processed at US poultry plants for indications of diseases or defects, visually examining the exterior, the inner surfaces of the body cavity, and the internal organs of every chicken carcass during processing operations. More recently, the USDA Food Safety and Inspection Service (FSIS) implemented the Hazard Analysis and Critical Control Point program throughout the country to ensure food safety and prevent food safety hazards in the inspection process for poultry, egg, and meat products and has also been testing the Hazard Analysis and Critical Control Point-based Inspection Models Project (USDA, 1996). Hazard Analysis and Critical Control Point-based Inspection Models Project food safety requirements include a zero tolerance performance standard for chickens with infectious conditions such as septicemia and toxemia; such birds must be removed from the processing line. To aid poultry plants in meeting government food safety regulations while maintaining their competitiveness and satisfying consumer demand, much research during the last decade has focused on the development of new inspection technologies such as automated computer vision inspection systems.

Previous studies have investigated both whole-carcass imaging and viscera-organ imaging methods for automated food safety inspection of poultry. Color imaging in red, green, blue color space for laboratory inspection of chicken spleens, hearts, and livers was found capable of identifying poultry disease conditions including leucosis, septicemia, airsacculitis, and ascites (Tao et al., 1998; Chao et al., 1999), but these methods require precise presentation of the visceral organs; most poultry processing line equipment is not suited for incorporating this type of viscera imaging. A 2-camera system for whole-carcass imaging at the 540- and 700-nm wavelengths was able to separate 90% of wholesome and unwholesome chickens on a 70-bird-per-minute (bpm) laboratory processing line but was unfeasible for higher speed processing (Park and Chen, 2000). Multivariate analysis of hyperspectral images (112 to 512 spectral channels) of whole chicken carcasses acquired using hyperspectral imaging systems under laboratory conditions was used for wavelength selection to develop 3- and 4-wavelength multispectral imaging systems for whole carcass inspection (Chao et al., 2002; Lawrence et al., 2003; Park et al., 2003). Despite effective wavelength selection, implementation with common-aperture multispectral camera hardware encountered significant difficulties during laboratory experiments. Adjusting exposure time for effective high-speed imaging of moving targets proved to be difficult for these multispectral systems due to their physical shutter mechanisms and the quantum efficiencies of the charge-coupled device cameras. An exposure time setting ideal for a shorter wavelength image often caused image saturation at a higher wavelength (Park et al., 2005; Yang et al., 2005, 2006). None of these studies moved beyond laboratory experiments to in-plant testing.

These problems can now be overcome by current electron-multiplying charge-coupled device (EMCCD)-based cameras, which have high quantum efficiencies and use rapid frame-transfer mechanisms and preoutput signal amplification via electron multiplication, significantly improving the signal-to-readout-noise ratio. In a line-scan imaging system, software controls eliminate the need for physical operation of a shutter and can operate the same EMCCD camera for either hyperspectral or multispectral imaging without any hardware adjustments. Consequently, wavelengths selected via hyperspectral analysis can be easily implemented for high-speed multispectral imaging of moving targets without the need for cross-system calibration.

American chicken plants now process over 8 billion birds annually. Federal regulations prohibit the sale of any birds showing signs of septicemia or toxemia, which typically occur for less than 0.2% of the birds processed. Septicemia is caused by the presence of pathogenic microorganisms or their toxins in the bloodstream, and toxemia is the result of toxins produced from cells at a localized infection or from the growth of microorganisms. Individual USDA inspectors conduct bird-by-bird inspections at maximum speeds of 35 bpm to remove all septicemic or toxemic chickens. The inspection process is subject to human variability, and the inspection speed restricts the maximum output possible for the processing plants while making inspectors prone to fatigue and repetitive injury problems. The objective of this study was to develop a multispectral image classification method using an algorithm based on fuzzy logic, implementing key wavelengths determined from hyperspectral image data collected online from a commercial processing line during in-plant testing of the hyperspectral-multispectral line-scan imaging system, for identifying systemically diseased chickens exhibiting septicemia or toxemia.

MATERIALS AND METHODS

Hyperspectral-Multispectral Line-Scan Imaging System

The imaging system consisted of a PhotonMAX 512b EMCCD camera (Princeton Instruments, Roper Scientific Inc., Trenton, NJ), an ImSpector V10 imaging spectrograph (Spectral Imaging Ltd., Oulu, Finland), and a pair of high-power, broad-spectrum white light-emitting-diode (LED) line lights (LL6212, Advanced Illumination Inc., Rochester, VT). The 16-bit camera had a back-illuminated EMCCD and a pixel read-out rate of 10 MHz. The current for the LED lights was set at 100 mA. With a 50-μm slit in front of the imaging spectrograph, each line scan (frame) acquired by the camera results in a spectrally contiguous series of up to 512 spectral line images.

The default line-scan image size of 512 ×512 pixels was reduced to 256 ×128 pixels by binning the pixels by 2 in the spatial dimension and by 4 in the spectral dimension. During spectral calibration using a mercury-neon lamp (Oriel Instruments, Stratford, CT), the intensities acquired from the first 20 and last 53 spectral channels were too low to be useful. Discarding these channels, the final line-scan image size was 256 (spatial) ×55 (spectral) pixels. To calibrate the relationship between spectral channels and wavelengths, spectral peaks were identified at specific channels corresponding to known mercury and neon reference wavelengths. Mercury peaks at 436 and 546 nm were found to correspond to the 8th and 25th channels, respectively, and neon peaks at 614, 640, 703, and 725 nm corresponded to the 35th, 38th, 48th, and 51st channels, respectively. The following second-order polynomial regression, in which λ= the wavelength in nm and nc = the spectral channel number, was calculated from the mercury and neon spectral peaks to calibrate the spectral axis:

 

(1)
\[{\lambda}\ =0.0034\ {\times}\ \mathit{n}_{\mathit{c}}^{2}\ 6.5362\ {\times}\ \mathit{n}_{\mathit{c}}\ +\ 382.74\]

The correlation coefficient of the linear regression between calibrated and expected wavelengths was 0.9995. From the wavelength calibration, the image spectrum ranged from 389 nm (the first channel) to 753 nm (the 55th channel), with an average bandwidth of 6.74 nm. The 160-mm linear field of view was translated into 256 spatial pixels, with each pixel representing an area of 0.625 ×0.625 mm2.

Figure 11 shows the line-scan imaging system set up to acquire images of chickens on a commercial evisceration line. The chicken shackles passed 914 mm in front of the camera lens and 267 mm in front of the plane of the 2 LED line lights, which were separated by 216 mm. The camera was set to acquire line-scan images using a 1-ms exposure time. The imaging system was able to acquire images at a speed of 400 line-scan images per second. A black acrylic background panel with a matte surface finish was mounted behind the shackles.

Hyperspectral Image Analysis

Using line-scan images W and D corresponding to Spectralon (Labsphere Inc., North Sutton, NH) and dark-field reference reflectances, respectively, that were collected each day before chicken imaging, flat-field correction was applied to line-scan chicken images used for hyperspectral image analysis. For each raw line-scan image I0, the pixel-based flat field correction was performed to obtain the relative reflectance for line-scan image I as follows:

 

(2)
\[\mathit{I}\ =\ \frac{\mathit{I}_{\mathit{0}}{-}\mathit{D}}{\mathit{W}{-}\mathit{D}}\]

The relative reflectance at 629 nm was used to create a mask for identifying chicken pixels in the line-scan images: pixels with relative reflectance values at 629 nm that were greater than or equal to 0.1 were assigned the value of 1 (chicken), and those with reflectance values less than 0.1 were assigned the value of zero (background). For the 55 spectral images of each line-scan, chicken pixels were identified using this mask image. Each set of 400 line-scan images (flat-field corrected and masked) was compiled to form a hyperspectral image cube of a whole chicken.

For chicken classification, determination of a suitable region of interest (ROI) within the chicken image and key wavelengths were required. For this end, an edge detection method was developed to locate the chicken edges and potential ROI boundaries; suitability for line-scan imaging applications was necessary for future use during online processing. Line scan by line scan, the algorithm detects the entry and exit of each bird from the field of view of the camera and determines the ROI for the carcass, as shown in Figure 22. With each line scanned, the relative reflectance at 629 nm is examined for each pixel within the uppermost 62.5 mm of the image (the top 100 pixels in this case), termed the carcass detection length in Figure 22. When the relative reflectance at 629 nm increases above 0.1 for any single pixel within the carcass detection length, the bird has entered the field of view. The detection algorithm examines only the uppermost 100 pixels to disregard wings and any eviscerated organs that may be positioned anomalously. After the first pixel with a reflectance greater than 0.1 at 629 nm is detected, the algorithm then monitors the subsequent scans as additional pixels within the detection length also begin showing relative reflectance greater than 0.1. Between the first detected pixel and the 100th pixel in the detection length, pixels below the first detected pixel will begin increasing in relative reflectance as the chicken continues to move across the field of view until there is a last scan with 1 (or several) pixel left that is located below the first detected pixel and is still within the detection length and whose reflectance has not increased above 0.1 but is found to increase above 0.1 in the next scan. This pixel, or the pixel in the center of a group of several pixels, is termed the turning point (TP), as shown in Figure 22, and represents the junction between the thigh and the belly on the leading edge of the bird. The line-scan images continue as the body of the chicken moves across the field of view, and in each of these line scans, the pixel at the same vertical coordinate as the TP will have increased above 0.1. However, once a line scan is acquired that shows the pixel at that coordinate with a relative reflectance below 0.1 at 629 nm, this pixel is termed the opposite TP (OTP), as shown in Figure 22, indicating that the main body of the bird has passed the field of view. Of the line scans that have been acquired for the chicken carcass up to this point, only those line scans containing a pixel at the same vertical coordinate as the TP and OTP with a reflectance greater than 0.1 will be used for analysis for chicken classification. After the OTP pixel is located, image acquisition can be stopped for this bird, because the remaining bird areas (such as the wingtip) are not needed for bird classification.

Within the line-scan images between the TP and OTP of each bird, only pixels including and below the vertical coordinate of the TP are considered for inclusion in the ROI to be used for analysis and bird classification, to eliminate the regions showing irregular shadows such as the legs, thighs, and wingtips, as shown by the contour map in Figure 33. Two values, m and n, were used to define the potential ROI boundary within this area. These values each specify a percentage of the pixels within the line-scan image, counted from the TP vertical coordinate down to the bottommost nonbackground pixel, as shown in Figure 33. Using a calibration set of images of wholesome and systemically diseased chickens, potential values of m and n were evaluated. Values of m from 10 to 50% were evaluated in combination with values of n from 60 to 90%. For each potential combination of m and n, the average wholesome pixel spectrum was calculated using pixels extracted from that ROI in every wholesome chicken line-scan image in the calibration set; the average systemically diseased pixel spectrum was similarly calculated using the systemically diseased chicken line-scan images. The difference spectrum between the average wholesome spectrum and average systemically diseased spectrum was calculated for each combination of m and n, and the minimum and maximum spectral differences were noted. The combination of m and n resulting in the highest spectral differences between wholesome and systemically diseased was selected as the ROI boundary values. Determination of key wavelengths was based on peaks in the difference spectrum for this ROI.

Multispectral Image Classification

The ROI and key wavelengths developed through hyperspectral image analysis were used to develop a classification algorithm based on fuzzy logic membership functions derived from fuzzy set theory (Zadeh, 1965). For algorithm development, the Fuzzy Logic Toolbox version 2.1.3 of MATLAB was used (MathWorks, Natick, MA).

To develop the fuzzy logic-based differentiation algorithm, a membership function was defined for each key wavelength, using the mean wholesome and systemically diseased intensity values and their SD. As seen in Figure 44, each membership function includes 2 fuzzy sets, 1 for systemically diseased chicken and 1 for wholesome chicken. For a given input intensity value, the degree of membership in the systemically diseased set was equal to 1 when the intensity input was equal to or less than the mean systemically diseased reflectance intensity minus 1 SD (Smstds). The degree of membership in the systemically diseased set was equal to 0 when the intensity input was equal to or greater than the mean systemically diseased reflectance intensity plus 1 SD (Sm+stds). For input values in between, the degree of membership decreased linearly from 1 to 0. The degree of membership in the wholesome set was equal to 0 when the intensity input was equal to or less than the mean wholesome reflectance intensity minus 1 SD (Wmstdw. The degree of membership in the wholesome set was equal to 1 when the intensity input was equal to or greater than the mean wholesome reflectance intensity plus 1 SD (Wm+stdw). For input values in between, the degree of membership increases linearly 0 to 1.

For each pixel in the ROI of each line-scan image, the membership functions were used to evaluate the key wavelength intensity value. For each key wavelength intensity value, 2 fuzzy membership values (s and w) in the systemically diseased and wholesome fuzzy sets were obtained. The fuzzy inference engine executes a min-max operation (Chao et al., 1999) to obtain a decision output (Do) for each pixel based on the n membership functions as follows, where n = the number of key wavelengths:

 

(3)
\[\mathit{D}_{o}\ =\ \mathit{max\ [min\ {\{}w}_{\mathit{l}}\ \mathit{...w}_{\mathit{n}}\mathit{,\ min\ {\{}s}_{\mathit{l}}\ \mathit{...s}_{\mathit{n}}\mathit{{\}}]}\]

For each pixel, the discrete decision output of Do was set to 1 when min {sl...sn} was greater than min {wl...wn}, indicating the existence of systemic disease; 0 when min {wl...wn} was greater than min {sl...sn}, indicating the nonexistence of systemic disease (i.e., the evidence of being wholesome); and 0.5 when the 2 were equal, indicating uncertainty of decision.

The discrete decision output values for all the ROI pixels of a compiled chicken image were then averaged, and the final decision for each bird was determined using a threshold value: for an average value above the threshold, the chicken was identified as being systemically diseased; otherwise, the bird was identified as being wholesome.

Experiment

Imaging control software was specifically developed in-house using Labview 8.0 (National Instruments Corp., Austin, TX) for the hyperspectral-multispectral line-scan imaging system. Hyperspectral image acquisition was conducted over two 1-wk periods at a commercial chicken processing plant. Chickens on an evisceration line moving at a speed of 70 bpm were identified as either wholesome or systemically diseased by an FSIS veterinarian. A set of 400 line-scan images was acquired for each chicken as it moved through the field of view of the camera. A calibration set containing 607 images of 543 wholesome chickens and 64 systemically diseased chickens was acquired during the first week. A set of 416 images of 381 wholesome chickens and 35 systemically diseased chickens was acquired during the second week. In addition, due to the low number of systemically diseased chickens encountered during the first week, 65 systemically diseased chickens were acquired during off-shift imaging in the second week. These 65 chickens were collected from bins of chickens that had been inspected and condemned by USDA inspectors. Following verification by the FSIS veterinarian for systemic disease, these birds were rehung onto the processing line for imaging during 15-min breaks between normal work shifts of the processing plant. After imaging, these birds were then immediately returned to the bins for USDA-condemned birds, and the shackles on the processing line were washed down. The testing set for this experiment consisted of all 481 images collected during the second week.

The 607 images in the calibration data set were used for hyperspectral data analysis to determine the ROI and key wavelengths and develop the fuzzy logic-based classification algorithm for classifying wholesome and systemically diseased chickens. The 481 images in the testing data set were used to validate the classification model.

RESULTS AND DISCUSSION

Hyperspectral Imaging and Analysis

Figure 55 shows example segmented compiled images of wholesome and systemically diseased chickens from the 1st, 7th, 37th, and 55th channels of the line-scan imaging system, which corresponded to the 389-, 429-, 629-, and 753-nm wavelengths. The line-scan imaging system was able to acquire high-quality short-wavelength images such as those from the 1st and 7th channels without saturating higher-wavelength images such as those from the 37th and 55th channels, due to the high quantum efficiency of the blue-sensitive back-lit charge-coupled device camera used in this system, unlike systems used in previous research (Yang et al., 2005, 2006). The line-scan imaging system had no difficulty in acquiring hyper- spectral images using 55 channels for birds moving at approximately 18 cm/s on the 70-bpm evisceration line. Each compiled image was composed of 400 line-scan images, containing 256 pixels from top to bottom. The wing-tip-to-wingtip width of each bird encompassed, on average, 275 line scans for wholesome birds and 250 line scans for systemically diseased birds, because systemically diseased birds tend to be slightly smaller than wholesome birds.

Figure 66 shows the range of the average spectral differences between wholesome and systemically diseased chickens in the calibration set that occurred for the ROI defined by a range of values for m and n. For each ROI, the spectral difference range includes the average spectral difference at each of the 55 wavelengths. The results indicated that when the value of n (the lower boundary of the ROI) decreased from 90 to 60% for a constant value of m (the upper boundary of the ROI), the spectral difference values increased. The increase in spectral difference values associated with decreasing the lower boundary was most significant for upper boundary values of 40 and 50%. When the value of m increased from 10 to 40% with a constant value of n, again, the spectral difference values increased. This increase in spectral difference values, associated with increasing the upper boundary within the 10 to 40% range, was most significant when the lower boundary was 60%. However, increasing the upper boundary m from 40 to 50% resulted in a slight decrease in the spectral difference values, suggesting a maximum in spectral difference occurring with a value of m from 40 to 50%. Consequently, the ROI boundaries m and n were set at 45 and 55%, respectively, resulting in the largest spectral difference values shown in Figure 66.

For the 45 to 55% ROI, Figure 77 shows the average spectra for wholesome and systemically diseased chicken pixels, with 1 SD envelope for each, from which the difference spectrum in Figure 88 was calculated. The difference spectrum shows 4 major peaks at 442, 501, 582, and 629 nm, which were selected as 4 key wavelengths for multi-spectral classification. The 629-nm wavelength was already determined to be useful for image masking. The other 3 wavelengths appear to match wavebands that were assigned by previous research to species of myoglobin whose presence in chicken meat reflects wholesome and unwholesome meat condition. All 3 species of myoglobin are present in chicken meat but interconvert through oxidation and degradation as the meat condition changes. Wholesome chicken meat shows more variations in oxymyoglobin and deoxymyoglobin, whereas unwholesome meat shows more variations in metmyoglobin. Through 2-dimensional correlation spectroscopy, analysis of chicken meat samples assigned wavebands at 485 and 500 nm to metmyoglobin, wavebands at 545 and 575 nm to oxymyoglobin, and a waveband at 440 nm to deoxymyoglobin (Liu et al., 2000; Liu and Chen, 2001).

Multispectral Classification

For multispectral classification, 1 fuzzy logic membership function was built for each of the 4 key wavelengths, using the mean and SD values shown in Table 11. Figure 99 shows the average fuzzy outputs for the 543 wholesome and 64 systemically diseased birds in the calibration data set. Using a threshold of 0.40, 90.6% of the wholesome birds and 93.8% of the systemically diseased birds were correctly identified. Only 4 systemically diseased chickens produced decision outputs below 0.40. Figure 1010 shows the average fuzzy outputs for 381 wholesome and 35 systemically diseased birds in the testing data set. Using the threshold of 0.40, 97.6% of the wholesome birds and 91.4% of the systemically diseased birds were correctly identified, with only 3 systemically diseased chickens misidentified. Figure 1111 shows the outputs for the 65 systemically diseased birds in the testing data set that were imaged off-shift. Using the threshold of 0.40, 100% of these birds were correctly identified. It should be noted that these systemically diseased birds may have been slightly dehydrated, because they were stored in bins of USDA-condemned birds during processing and later hung on the processing line for imaging. The combined classification accuracy for all systemically diseased birds in the testing set was 96%.

Applications

The fuzzy logic-based multispectral classification method developed in this study can be easily implemented for online multispectral chicken inspection using the line-scan imaging system that was used to acquire the hyperspectral data for method development. Furthermore, multispectral inspection can be implemented for processing lines running at even higher speeds, because the imaging system can perform more quickly for acquisition and processing of data at only 4 wavelengths, compared with the 55 wavelengths used for analysis.

The capacity for a higher operating speed would allow this system to be used on poultry kill lines, which may run at speeds of 140 or 180 bpm, as prescreening tools to prevent most systemically diseased birds from being transferred to the evisceration line for further processing and for inspection by FSIS inspectors. For this purpose, the classification threshold can be adjusted to ensure near-100% accuracy in identifying systemically diseased birds for rejection, with a low rate of erroneously rejected wholesome birds. The multispectral classification system would require implementation in conjunction with commercial poultry processing equipment, already available, that can automatically divert rejected birds (identified as systemically diseased by the imaging system) while feeding the wholesome birds to the evisceration line. Reinspection of the birds on the lower-volume rejection line by inspection personnel to catch misidentified wholesome birds would minimize economic loss for processors. The evisceration line would be fed only wholesome birds requiring no additional inspection, allowing for higher speed processing while minimizing food safety risks and cross-contamination of equipment. Figure 1212 shows the classification accuracies for wholesome and systemically diseased birds in the testing data set for threshold values from 0.10 to 0.90. By adjusting the classification threshold to 0.30, for example, 100% of systemically diseased birds would be identified for rejection, with approximately 11% of wholesome birds erroneously rejected as well.

Paw harvesting (i.e., the processing of chicken feet) would be another application for the multispectral classification system. Feet are removed from each chicken carcass while it is on the kill line, but FSIS food safety inspection of the chicken carcass does not occur until much further downstream on the evisceration line. The time between inspection of a particular carcass and the removal of its feet may be as much as 15 min, and food safety regulations indicate that feet of condemned carcasses cannot be sold. Consequently, processors generally either 1) store batches of removed feet in hoppers and wait until the corresponding carcasses have passed inspection, with the condition that an entire hopper may be discarded if any of the carcasses are condemned, or 2) maintain tracking systems with extra lengths of kill line shackles that temporarily continue to hold the removed feet of each carcass while the carcass is transferred to the evisceration line so that when the carcass reaches inspection, carcass rejection can be tracked and used to trigger the rejection of its corresponding feet and no others. By implementing the multispectral classification system on the kill line for 100% accuracy in identifying systemically diseased birds, the chicken feet of these birds (and a small number of misidentified wholesome birds) could be immediately rejected, allowing harvesting of wholesome feet with lower economic losses compared with the hopper system or lower equipment requirements compared with the track-and-hold system.

In conclusion, the hyperspectral-multispectral line-scan imaging system can be used for automated online inspection of chicken carcasses for the detection of systemically diseased birds on high-speed processing lines. In-plant testing demonstrated an imaging speed of 400 line-scan images per second, which can easily handle high-speed processing operations, and implementation of the imaging system on processing lines can be adapted to suit specific screening-processing goals by adjusting the classification threshold used for separating wholesome and systemically diseased birds. Applications for the system include automated food safety inspection of birds for prescreening that increases processing efficiency by eliminating systemically diseased birds from evisceration lines, which consequently can be operated more quickly, and for kill line paw-harvesting operations that can minimize the loss of wholesome paws with lower equipment requirements compared with current paw-harvesting operations.

Table 1.

Mean and SD values for the membership functions constructed from relative reflectance intensity I at the 442-, 501-, 582-, and 629-nm key wavelengths

 Wholesome Systemically diseased 
Membership function inputs Mean SD Mean SD 
I442 0.3668 0.1571 0.2491 0.1072 
I501 0.4238 0.1536 0.3003 0.1078 
I582 0.4095 0.1406 0.2913 0.1045 
I629 0.5225 0.1559 0.4059 0.1133 
 Wholesome Systemically diseased 
Membership function inputs Mean SD Mean SD 
I442 0.3668 0.1571 0.2491 0.1072 
I501 0.4238 0.1536 0.3003 0.1078 
I582 0.4095 0.1406 0.2913 0.1045 
I629 0.5225 0.1559 0.4059 0.1133 
Figure 1.

Hyperspectral-multispectral line-scan imaging system positioned for chicken imaging on a commercial evisceration line moving at a speed of 70 birds per minute.

Figure 1.

Hyperspectral-multispectral line-scan imaging system positioned for chicken imaging on a commercial evisceration line moving at a speed of 70 birds per minute.

Figure 2.

Carcass detection length, turning Point (TP), and opposite turning point (OTP).

Figure 2.

Carcass detection length, turning Point (TP), and opposite turning point (OTP).

Figure 3.

Contour images of a wholesome chicken (left) and a systemically diseased chicken (right), compiled from individual line-scan images. For each line-scan image, the upper and lower boundaries of the region of interest within the 100% length were defined by values of m and n, respectively.

Figure 3.

Contour images of a wholesome chicken (left) and a systemically diseased chicken (right), compiled from individual line-scan images. For each line-scan image, the upper and lower boundaries of the region of interest within the 100% length were defined by values of m and n, respectively.

Figure 4.

Construction of fuzzy logic membership function for each key wavelength, for wholesome and systemically diseased chicken, based on mean and 1 SD values for reflectance intensity. Sm±stds = average systemically diseased spectrum with 1 SD envelope; Wm±stdw = average wholesome spectrum with 1 SD envelope. fi = intensity at key wavelength.

Figure 4.

Construction of fuzzy logic membership function for each key wavelength, for wholesome and systemically diseased chicken, based on mean and 1 SD values for reflectance intensity. Sm±stds = average systemically diseased spectrum with 1 SD envelope; Wm±stdw = average wholesome spectrum with 1 SD envelope. fi = intensity at key wavelength.

Figure 5.

Images of 1 wholesome chicken and 1 systemically diseased chicken at the wavelengths of 389, 429, 629, and 753 nm, compiled from 400 line-scan images for each bird. Each line-scan image contained 256 pixels.

Figure 5.

Images of 1 wholesome chicken and 1 systemically diseased chicken at the wavelengths of 389, 429, 629, and 753 nm, compiled from 400 line-scan images for each bird. Each line-scan image contained 256 pixels.

Figure 6.

The range of intensity difference between wholesome and systemically diseased chickens, for each region of interest as defined by values of m and n.

Figure 6.

The range of intensity difference between wholesome and systemically diseased chickens, for each region of interest as defined by values of m and n.

Figure 7.

Average wholesome spectrum (Wm) and average systemically diseased spectrum (Sm), with 1 SD envelope (Wm±stdw, Sm±stds), for the region of interest defined by m = 45% and n = 55%.

Figure 7.

Average wholesome spectrum (Wm) and average systemically diseased spectrum (Sm), with 1 SD envelope (Wm±stdw, Sm±stds), for the region of interest defined by m = 45% and n = 55%.

Figure 8.

Difference spectrum calculated from the average wholesome and average systemically diseased spectra. Four key wavelengths were selected at 442, 501, 582, and 629 nm.

Figure 8.

Difference spectrum calculated from the average wholesome and average systemically diseased spectra. Four key wavelengths were selected at 442, 501, 582, and 629 nm.

Figure 9.

Fuzzy output decision values for the images of 543 wholesome and 64 systemically diseased chickens in the calibration data set, plotted against total number of region of interest pixels for each bird.

Figure 9.

Fuzzy output decision values for the images of 543 wholesome and 64 systemically diseased chickens in the calibration data set, plotted against total number of region of interest pixels for each bird.

Figure 10.

Fuzzy output decision values for the images of 381 wholesome and 35 systemically diseased chickens in the testing data set (excluding off-shift imaged birds), plotted against total number of region of interest pixels for each bird.

Figure 10.

Fuzzy output decision values for the images of 381 wholesome and 35 systemically diseased chickens in the testing data set (excluding off-shift imaged birds), plotted against total number of region of interest pixels for each bird.

Figure 11.

Fuzzy output decision values for the images of 65 systemically diseased chickens imaged off-shift in the testing data set, plotted against total number of region of interest pixels for each bird.

Figure 11.

Fuzzy output decision values for the images of 65 systemically diseased chickens imaged off-shift in the testing data set, plotted against total number of region of interest pixels for each bird.

Figure 12.

The differentiation accuracy by various threshold values to differentiate wholesome and systemically diseased chickens in the testing data set by fuzzy output decision values.

Figure 12.

The differentiation accuracy by various threshold values to differentiate wholesome and systemically diseased chickens in the testing data set by fuzzy output decision values.

1
Mention of company or trade names is for descriptive purposes only and does not imply endorsement or approval by the USDA to the exclusion of other products that may be suitable.

REFERENCES

Chao
,
K.
, Y. R. Chen, H. L. Early, and B. Park.
1999
. Color image classification systems for poultry viscera inspection.
Appl. Eng. Agric.
 
15
:
363
–369.
Chao
,
K.
, P. M. Mehl, and Y. R. Chen.
2002
. Use of hyper- and multi-spectral imaging for detection of chicken skin tumors.
Appl. Eng. Agric.
 
18
:
78
–84.
Lawrence
,
K. C.
, B. Park, W. R. Windham, and C. Mao.
2003
. Calibration of a pushbroom hyperspectral imaging system for agricultural inspection.
Trans. ASAE
 
46
:
513
–521.
Liu
,
Y.
, and Y. R. Chen.
2001
. Analysis of visible reflectance spectra of stored, cooked and diseased chicken meats.
Meat Sci.
 
58
:
395
–401.
Liu
,
Y.
, Y. R. Chen, and Y. Ozaki.
2000
. Characterization of visible spectral intensity variations of wholesome and unwholesome chicken meats with two-dimensional correction spectroscopy.
Appl. Spectrosc.
 
54
:
587
–594.
Park
,
B.
, and Y. R. Chen.
2000
. Real-time dual-wavelength image processing for poultry safety inspection.
J. Food Proc. Eng.
 
23
:
329
–351.
Park
,
B.
, K. C. Lawrence, W. R. Windham, and R. J. Buhr.
2003
. Hyperspectral imaging for detecting fecal and ingesta contaminants on poultry carcasses.
Trans. ASAE
 
45
:
2017
–2026.
Park
,
B.
, K. C. Lawrence, W. R. Windham, and D. P. Smith.
2005
. Multispectral imaging system for fecal and ingesta detection on poultry carcasses.
J. Food Proc. Eng.
 
27
:
311
–327.
Tao
,
Y.
, J. Shao, K. Skeeles, and Y. R. Chen.
1998
. Detection of eviscerated poultry spleen enlargement by machine vision.
SPIE
 
3544
:
138
–145.
USDA.
1996
. Pathogen Reduction: Hazard Analysis and Critical Control Point (HACCP) Systems, final rule. 9CFR part 304.
Fed. Regist.
 
61
:
38805
–38989.
Yang
,
C. C.
, K. Chao, and Y. R. Chen.
2005
. Development of multispectral imaging processing algorithms for identification of wholesome, septicemic, and inflammatory process chickens.
J. Food Eng.
 
69
:
225
–234.
Yang
,
C. C.
, K. Chao, Y. R. Chen, M. S. Kim, and H. L. Early.
2006
. Simple multispectral image analysis for systemically diseased chicken identification.
Trans ASABE.
 
49
:
245
–257.
Zadeh
,
L. A.
1965
. Fuzzy sets.
Inf. Controls.
 
8
:
338
–353.