Skip to main content
ARS Home » Research » Publications at this Location » Publication #184319
Title: Supervised classification algorithms for poultry hyperspectral image analysis

Author
item Park, Bosoon
item Lawrence, Kurt
item Windham, William
item Smith, Douglas

Submitted to: Near Infrared Spectroscopy International Conference Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 6/30/2005
Publication Date: 3/1/2007
Citation: Park, B., Lawrence, K.C., Windham, W.R., Smith, D.P. 2007. Supervised classification algorithms for poultry hyperspectral image analysis. Near Infrared Spectroscopy International Conference Proceedings. p. 239-244.

Interpretive Summary: Food safety has become an important issue for public health, because reduction in the potential health risks to consumers from human pathogens in food is the most important food safety issue and public concern. Hyperspectral imaging technique demonstrated potential for the food safety inspection, especially poultry fecal detection. Classification of hyperspectral imagery was developed to identify the type and source of various fecal contaminants to improve federal poultry safety program for Hazard Analysis Critical Control Point (HACCP). Several different classification methods were investigated to determine the optimum method having best performance to classify surface contaminants on broiler carcasses. A hyperspectral imaging system with selected classifier can improve federal poultry safety inspection program, incorporating with scientific testing and systematic prevention of contamination.

Technical Abstract: A hyperspectral imaging system with optimum classifiers enables us to identify the type and source of various contaminants (duodenum, ceca, colon, and ingesta) to determine critical control point for science-based federal poultry safety inspection program. The classification accuracies varied from 62.9% to 92.3%, depending on classification method as well as the type of diets. For the selection of optimum classification methods, the parameter of each classifier needs to be determined. It was found that optimum threshold value for Mahalanobis distance could decrease false positives and increase classification accuracy. Since the hyperspectral imagery for this study contained 512 bands, which could be redundant, the optimum number of bands needs to be selected to minimize data processing time and maximize system performance.