Submitted to: Journal of Food Engineering
Publication Type: Peer reviewed journal
Publication Acceptance Date: 7/12/2006
Publication Date: 1/1/2007
Citation: Chao, K., Chen, Y.R., Ding, F., Yang, C., Chan, D.E. 2007. Development of two-band color-mixing technique for identification of broiler carcass conditions. Journal of Food Engineering. 80(1):276-283. Interpretive Summary: FSIS has completed the transformation of its traditional inspection system to a Hazard-Analysis-and-Critical-Control-Point (HACCP) inspection system. With more focus on HACCP and HIMP (HACCP-based inspections models project), the FSIS has placed more of the burden of inspection responsibility on the processors. In essence, processors assume the responsibility for inspection, and the regulatory agency performs oversight and verification to ensure standards are met. HACCP or HIMP helps meet consumer demand for safe, high quality food; however, consumer demand for more food increases the need for, and pressure on, inspectors. Consequently, the development of accurate, rapid, and non-invasive technologies appropriate for operation on high-speed processing lines is of great importance for the poultry industry. The overall objective of this paper was to investigate the two-band color-mixing technique for automated poultry inspection. All pairwise combinations of 10-nm wavebands in the region of 416 nm - 715 nm were examined for differentiating between chicken conditions using the two-band color-mixing technique, and the waveband pair of (453 nm, 589 nm) was found most suitable for identifying the chicken conditions. Significant color differences between wholesome, systemically diseased, and cadaver chicken samples, as indicated by L*, u*, and v* values, were found using one-way analysis of variance. Color difference values calculated from L*, u*, and v* values were used as inputs for decision-tree classification models. For a validation data set containing 48 wholesome, 34 systemically diseased, and 22 cadaver chicken samples, the classification accuracies achieved were 95.8% for wholesome, 95.5% for systemically diseased, and 100% for cadaver. For an independent testing set of 55 wholesome, 32 systemically diseased, and 18 cadaver chicken samples, the classification accuracies were 94.6%, 100%, and 90.6%, respectively. For model testing using 10-fold cross-validation, classification accuracies achieved were 98.1% for wholesome, 97.5% for systemically diseased, and 93.9% for cadaver. These results show that the two-band color-mixing technique shows potential for use in addressing public health concerns, as only 1 unwholesome (a cadaver sample) bird was misclassified as wholesome. This information is useful to the Food Safety and Inspection Service (FSIS), and poultry equipment and processing plants.
Technical Abstract: The development of accurate, rapid, and non-invasive inspection technologies are needed to help poultry processors meet food safety regulations and rising consumer demand while increasing productivity and economic competitiveness. This paper reports on a novel two narrow-band color-mixing technique for identification of broiler chicken carcass conditions. Spectra were collected for samples taken from 103 wholesome chicken carcasses, 66 systemically diseased chicken carcasses, and 40 cadaver chicken carcasses using a photodiode array spectrophotometer system. Waveband pairs in the range of 416 nm - 715 nm were evaluated for identifying chicken conditions using the two-band color-mixing technique, and the pair of (453 nm, 589 nm) was selected based on color difference index calculations in CIELUV color space. Significant differences in the color characteristics of wholesome, systemically diseased, and cadaver chicken conditions, based on color-mixing using the two selected wavebands, were confirmed by one-way analysis of variance. Decision-tree classification models using the calculated color difference indexes were evaluated first by using the spectral data divided into a validation set and a testing set, and second by 10-fold cross-validation of the entire data set. Classification accuracies achieved for the wholesome, systemically diseased, and cadaver samples were 95.8%, 95.5%, and 100%, respectively, for the validation set; 94.6%, 100%, and 90.6%, respectively, for the testing set; and 98.1%, 97.5%, and 93.9%, respectively, when using 10-fold cross-validation.