SYSTEMICALLY DISEASED CHICKEN IDENTIFICATION USING MULTISPECTRAL IMAGES AND REGION OF INTEREST ANALYSIS

Authors: YANG, CHUN-CHIEH - VIS.SCI.,U.KY; Chao, Kuanglin - Kevin Chao; Chen, Yud; EARLY, HOWARD - USDA,FSIS,WASH.,DC

Submitted to: International Society for Optical Engineering
Publication Type: Proceedings
Publication Acceptance Date: 2/20/2005
Publication Date: 6/1/2005
Citation: Yang, C.C., Chao, K., Chen, Y.R. 2005. Development of multispectral imaging processing algorithms for food safety inspection on poultry carcasses. J. of Food Engineering. 69(2):225-234.

Interpretive Summary: In this study, simple multispectral image processing and classification method was developed to differentiate wholesome and systemically diseased chickens. Because of significant color differences between wholesome and systemically diseased chickens at the wavelength of 540 nm, the interference filter was selected for two different multispectral imaging systems. An algorithm was developed to find the ROI on the multispectral images from different imaging systems. It is essential that the same Spectralon reference target was used in both imaging systems to acquire calibration reference images prior to collection of chicken images while the camera lens was covered at both imaging systems to acquire the dark reference image. This step was to make images from different imaging systems comparable to each other. With the average relative reflectance intensity as the input, classification thresholds for identifying wholesome and systemically diseased chickens were determined using a C&RT decision tree algorithm. Using the first image batch to generate threshold and applying it to the second image batch, the classification accuracies for the first image batch were 100% for wholesome chickens and 97.2% for systemically diseased chickens, and 99.7% for wholesome chickens and 93.5% for systemically diseased chickens for the second image batch. The generated threshold was 0.2572. Using the second image batch to generate threshold and applying it to the first image batch, the decision tree obtained the accuracies of 96.4% for wholesome chickens and 100% for systemically diseased chickens for the second image batch, and 95.7% for wholesome chickens and 97.7% for systemically diseased chickens for the first image batch. The threshold was 0.2626. This classification method, using the simple calculation of average relative reflectance intensity, showed a significant potential for testing in an automated on-line multispectral inspection system.

Technical Abstract: A simple multispectral classification method for the identification of systemically diseased chickens was developed and tested between two different imaging systems. An image processing algorithm was developed to define and locate the region of interest (ROI) as classification areas on the image. The average intensity was calculated for each classification area of the chicken image. A decision tree algorithm was used to determine threshold values for each classification areas. The wavelength of 540 nm was used for image differentiation purpose. There were 164 wholesome and 176 systemically diseased chicken images collected using the first imaging system, and 332 wholesome and 318 systemically diseased chicken images taken by the second imaging system. The differentiation thresholds, generated by the decision tree method, based on the images from the first imaging system were applied to the images from the second imaging system, and vice versa. The accuracy from evaluation was 95.7% for wholesome and 97.7% of systemically diseased chickens for the first image batch, and 99.7% for wholesome and 93.5% for systemically diseased chickens for the second image batch. The result showed that using single wavelength and threshold, this simple classification method can be used in automated on-line applications for chicken inspection.