Bosoon Park

Comparison between visible/NIR spectroscopy and hyperspectral imaging for detecting surface contaminants on poultry carcasses

The U. S. Department of Agriculture, Agricultural Research Service has been developing a method and system to detect fecal contamination on processed poultry carcasses with hyperspectral and multispectral imaging systems. The patented method utilizes a three step approach to contaminant detection. Spectra of homogenous samples of feces, ingesta (undigested food particles), and skin were first collected. Then those spectra were evaluated with multivariate analysis techniques to identify significant wavelength regions for further analysis. Hyperspectral data were then collected on contaminated poultry carcasses and information learned from the spectroscopic data was used to aide in hyperspectral data analysis. Finally, the results of the hyperspectral data were used to identify a few optimum wavelengths for use in a real-time multispectral imaging system. In this work, two techniques for developing spectral datasets and algorithms for classifying surface contaminants on poultry carcasses were explored. The first consisted of a scanning monochrometer that measured the average spectra of uncontaminated breast skin and fecal and ingesta contaminants. The second technique used regions of interest (ROI) from a hyperspectral image to collect spatially averaged spectra. Comparison of the spectra from each instrument showed variations in the spectra collected from similar samples. There was an offset of absorption values between the two instruments and the hyperspectral imaging system had better resolution at higher absorption wavelengths. Although both systems were calibrated prior to measuring, there was also a slight shift in absorption peaks between the two systems. Both techniques were able to classify contaminated skin from uncontaminated skin in a full cross-validated test set with better than 99% accuracy. However, when the classification model developed from the monochrometer spectra was applied to whole-carcass hyperspectral images, numerous common carcass features, such as exposed meat and wing-shadowed skin, were wrongly identified as false positives. Since spectra of entire poultry carcasses were available in the original hyperspectral dataset, the hyperspectral ROI technique allowed researchers to easily add the spectra of these false positives to the calibration dataset. New partial least squares regression models with meat and skin shadow spectra resulted in different principal component loadings and improved classification models. The classification model with the combined ROI spectra from skin, feces, ingesta, meat, and skin shadows gave a classification accuracy of 99.5%. When this model was compared to the original model developed from the monochrometer dataset on a few hyperspectral images of contaminated carcasses, fewer false positives were classified with the hyperspectral ROI model without sacrificing the accuracy of contaminant detection. Further research must be done to fully characterize the accuracy of the model.

Integration of visible/NIR spectroscopy and multispectral imaging for poultry carcass inspection

Abstract An integrated system which consisted of a visible/near-infrared (NIR) spectroscopic subsystem and an intensified multispectral imaging subsystem was tested for its accuracy in separating abnormal (unwholesome) from normal poultry carcasses. The spectroscopic subsystem measured reflectance spectra of the poultry carcasses at wavelengths from 471 to 965 nm. For the multispectral imaging subsystem, the gray-level intensity of whole carcasses was measured using six different optical filters of 542, 571, 641, 700, 720 and 847 nm wavelengths. The results of the test showed that, with the integrated system, there were no abnormal carcasses being misclassified as normal carcasses. When individual subsystem was used for classification, the error of the spectroscopic subsystem was 2.6% and that of the multispectral imaging subsystem was 3.9%. Thus, the integrated system could be used for separating carcasses into normal and abnormal streams. With perfect selection of normal carcasses in the normal carcass stream, the inspector needs only to inspect the abnormal carcass stream. Thus, the through-put of carcass of the processing line per inspector could be greatly increased.

The effect of regions of interest and spectral pre-processing on the detection of non-0157 Shiga-toxin producing Escherichia coli serogroups on agar media by hyperspectral imaging

Foodborne infection caused by Shiga toxin-producing Escherichia coli (STEC) is a major worldwide health concern. The best known and highly virulent STEC serogroup is E. coli 0157:H7, which can be easily identified when cultured on sorbitol-MacConkey (SMAC) agar. Recently, six non-0157 STEC serogroups have been found to cause human illnesses. These non-0157 serogroups ferment sorbital and form pink colonies; therefore SMAC agar cannot be used to differentiate non-0157 serogroups from each other and other flora growing on the plate. This study investigated the potential of visible and near infrared hyperspectral imaging and chemometrics to spectrally differentiate six representative non-0517 STEC serogroups (026, 045, 0103, 0111, 0121 and 0145) grown as spots on Rainbow agar media. Mahalanobis distance classifiers were developed with spectra obtained from ground truth regions of interest (ROIs) of each serogroup colony. The ROIs were selected as a doughnut-like open-ellipse to only include the leading edge of growth and as a closed-ellipse covering the entire colony. For each ROI type, the Mahalanobis distance classifiers were developed with log (1/Reflectance), first derivative and standard normal variate and detrending (SNVD) pre-processing treatments. Serogroups 045 and 0121 were consistently classified over 98% accurate, regard less of the classification algorithm used. The lowest classification accuracies were from classifiers developed with only log (1/R) ROI spectra. First derivative and SNVD spectra helped to increase the detection accuracies of the other serogroups. The classification accuracy for serogroups 026, 0111, 0103 and 0145 with the closed-ellipse and open-ellipse classification algorithms showed varying results from 8% to 87% and 57% to 100%, respectively. The lower accuracies with closed ellipse spectra were due to greater spectral variation in the centre pixels on a per-pixel basis. Practical implications of this study are the demonstrated potential of hyperspectral imaging for presumptive-positive screening of non-0157 serogroups on Rainbow agar and the extensibility of the developed sampling methods and classification models for future research to identify the target bacteria in the presence of background flora grown on spread plates.

Hyperspectral imaging for food processing automation

This paper presents the research results that demonstrates hyperspectral imaging could be used effectively for detecting feces (from duodenum, ceca, and colon) and ingesta on the surface of poultry carcasses, and potential application for real-time, on-line processing of poultry for automatic safety inspection. The hyperspectral imaging system included a line scan camera with prism-grating-prism spectrograph, fiber optic line lighting, motorized lens control, and hyperspectral image processing software. Hyperspectral image processing algorithms, specifically band ratio of dual-wavelength (565/517) images and thresholding were effective on the identification of fecal and ingesta contamination of poultry carcasses. A multispectral imaging system including a common aperture camera with three optical trim filters (515.4 nm with 8.6- nm FWHM), 566.4 nm with 8.8-nm FWHM, and 631 nm with 10.2-nm FWHM), which were selected and validated by a hyperspectral imaging system, was developed for a real-time, on-line application. A total image processing time required to perform the current multispectral images captured by a common aperture camera was approximately 251 msec or 3.99 frames/sec. A preliminary test shows that the accuracy of real-time multispectral imaging system to detect feces and ingesta on corn/soybean fed poultry carcasses was 96%. However, many false positive spots that cause system errors were also detected.

Hybrid image processing for robust extraction of lean tissue on beef cut surfaces

Abstract A hybrid image processing system which automatically distinguishes lean tissues in the image of a complex beef cut surface and generates the lean tissue contour has been developed. Because of the inhomogeneous distribution and fuzzy pattern of fat and lean tissues on the beef cut, conventional image segmentation and contour generation algorithms suffer from a heavy computing requirement, algorithm complexity and poor robustness. The proposed system utilizes an artificial neural network to enhance the robustness of processing. The system is composed of pre-network, network, and post-network processing stages. At the pre-network stage, gray level images of beef cuts were segmented and resized to be adequate to the network input. Features such as fat and bone were enhanced and the enhanced input image was converted to a grid pattern image, whose grid was formed as 4 × 4 pixel size. At the network stage, the normalized gray value of each grid image was taken as the network input. The pre-trained network generated the grid image output of the isolated lean tissue. A sequence of post-network processing was conducted to obtain the detailed contour of the lean tissue. A training scheme of the network and the separating performance were presented and analyzed. The developed hybrid system showed the feasibility of the human-like robust object segmentation and contour generation for the complex, fuzzy and irregular image.

Analysis of reflectance spectra from hyperspectral images of poultry carcasses for fecal and ingesta detection

Identification and separation of poultry carcasses contaminated by feces and/or crop ingesta are very important to protect the consumer from a potential source of food poisoning. A transportable hyperspectral imaging system was developed to detect fecal and ingesta contamination on the surface of poultry carcasses. Detection algorithms used with the imaging system were developed from visible/near infrared monochromator spectra and with contaminates from birds fed a corn/soybean meal diet. The objectives of this study were to investigate using regions of interest reflectance spectra from hyperspectral images to determine optimal wavelengths for fecal detection algorithms from images of birds fed corn, wheat and milo diets. Spectral and spatial data between 400 and 900 nm with a 1.0 nm spectral resolution were acquired from uncontaminated and fecal and ingesta contaminated poultry carcasses. Regions of interest (ROIs) were defined for fecal and ingesta contaminated and uncontaminated skin (i.e. breast, thigh, and wing). Average reflectance spectra of the ROIs were extracted for analysis. Reflectance spectra of contaminants and uncontaminated skin differed. Spectral data pre-processing treatments with a single-term, linear regression program to select wavelengths for optimum calibration coefficients to detect contamination were developed. Fecal and ingesta detection models, specifically a quotient of 2 and/or 3-wavelengths was 100% successful in classification of contaminates.

Classification of hyperspectral imagery for identifying fecal and ingesta contaminants

This paper presents the research results of the performance of classification methods for hyperspectral poultry imagery to identify fecal and ingesta contaminants on the surface of broiler carcasses. A pushbroom line-scan hyperspectral imager was used to acquire hyperspectral data with 512 narrow bands covered from 400 to 900 nm wavelengths. Three different feces from digestive tracts (duodenum, ceca, colon), and ingesta were used as contaminants. These contaminants were collected from the broiler carcasses fed by corn, milo, and wheat with soybean meals. For the selection of optimum classifier, various widely used supervised classification methods (parallelepiped, minimum distance, Mahalanobis distance, maximum likelihood, spectral angle mapper, and binary coding) were investigated. The classification accuracies ranged from 62.94% to 92.27%. The highest classification accuracy for identifying contaminants for corn fed carcasses was 92.27% with spectral angle mapper classifier. While, the classification accuracy was 82.02% with maximum likelihood method for milo fed carcasses and 91.16% accuracy was obtained for wheat fed carcasses when same classification method was used. The mean classification accuracy obtained in this study for classifying fecal and ingesta contaminants was 90.21%.

Egg embryo development detection with hyperspectral imaging

In the U. S. egg industry, anywhere from 130 million to over one billion infertile eggs are incubated each year. Some of these infertile eggs explode in the hatching cabinet and can potentially spread molds or bacteria to all the eggs in the cabinet. A method to detect the embryo development of incubated eggs was developed. Twelve brown-shell hatching eggs from two replicates (n=24) were incubated and imaged to identify embryo development. A hyperspectral imaging system was used to collect transmission images from 420 to 840 nm of brown-shell eggs positioned with the air cell vertical and normal to the camera lens. Raw transmission images from about 400 to 900 nm were collected for every egg on days 0, 1, 2, and 3 of incubation. A total of 96 images were collected and eggs were broken out on day 6 to determine fertility. After breakout, all eggs were found to be fertile. Therefore, this paper presents results for egg embryo development, not fertility. The original hyperspectral data and spectral means for each egg were both used to create embryo development models. With the hyperspectral data range reduced to about 500 to 700 nm, a minimum noise fraction transformation was used, along with a Mahalanobis Distance classification model, to predict development. Days 2 and 3 were all correctly classified (100%), while day 0 and day 1 were classified at 95.8% and 91.7%, respectively. Alternatively, the mean spectra from each egg were used to develop a partial least squares regression (PLSR) model. First, a PLSR model was developed with all eggs and all days. The data were multiplicative scatter corrected, spectrally smoothed, and the wavelength range was reduced to 539 - 770 nm. With a one-out cross validation, all eggs for all days were correctly classified (100%). Second, a PLSR model was developed with data from day 0 and day 3, and the model was validated with data from day 1 and 2. For day 1, 22 of 24 eggs were correctly classified (91.7%) and for day 2, all eggs were correctly classified (100%). Although the results are based on relatively small sample sizes, they are encouraging. However, larger sample sizes, from multiple flocks, will be needed to fully validate and verify these models. Additionally, future experiments must also include non-fertile eggs so the fertile / non-fertile effect can be determined.

Improved imaging system for fecal detection

The Agricultural Research Service (ARS) has developed a hyperspectral imaging system to detect fecal contaminants on poultry carcasses. The system operates from about 400 to 1000 nm, but only a few wavelengths are used in a real-time multispectral system. ARS has reported that the ratio of reflectance images at 565 nm and 517 nm was able to identify fecal contaminants. However, this ratio alone also misclassified numerous non-fecal carcass features (false positives). Recent modifications to the system, including improved lighting, new camera, new spectrograph, and a new algorithm with an additional wavelength, have increased fecal detection accuracy while reducing the number of false positives. The new system was used to collect hyperspectral data on 56 stationary poultry carcasses. Carcasses were contaminated with both large and small spots of feces from the duodenum, ceca, and colon, and ingesta from the crop. A total of 1030 contaminants were applied to the carcasses. The new algorithm correctly identified over 99% of the contaminants with only 25 false positives. About a quarter of the carcasses had at least one false positive.

Hyperspectral imaging using RGB color for foodborne pathogen detection

Abstract. This paper reports the development of a spectral reconstruction technique for predicting hyperspectral images from RGB color images and classifying food-borne pathogens in agar plates using reconstructed hyperspectral images. The six representative non-O157 Shiga-toxin producing Escherichia coli (STEC) serogroups (O26, O45, O103, O111, O121, and O145) grown on Rainbow agar plates were used for the study. A line-scan pushbroom hyperspectral imaging spectrometer was used to scan full reflectance spectra of pure non-O157 STEC cultures in the visible and near-infrared spectral range from 400 to 1000 nm. RGB color images were generated by simulation from hyperspectral images. Polynomial multivariate least-squares regression analysis was used to reconstruct hyperspectral images from RGB color images. The mean R-squared value for hyperspectral image reconstruction was ∼0.98 in the spectral range between 400 and 700 nm for linear, quadratic, and cubic polynomial regression models. The accuracy of the hyperspectral image classification algorithm based on k-nearest neighbors algorithm of principal component scores was validated to be 92% with the test set (99% with the original hyperspectral images). The results of the study suggested that color-based hyperspectral imaging would be feasible without much loss of prediction accuracy compared to true hyperspectral imaging.