Submitted to: ASAE Annual International Meeting
Publication Type: Proceedings
Publication Acceptance Date: July 30, 2002
Publication Date: July 30, 2002
Citation: Lawrence, K.C., Park, B., Smith, D.P., Windham, W.R., Feldner, P.W. 2002. Hyperspectral system calibration for improved contaminant detection on poultry carcasses. American Society of Agricultural Engineers Annual International Meeting.
Interpretive Summary: An imaging camera system was developed to detect contamination on poultry carcasses. The imaging system is known as a hyperspectral imaging system, which can measure the light intensity over a range of wavelengths for every dot, or pixel, of a digital image. Since this camera was a research tool, a calibration protocol was developed earlier. This paper evaluates the effects camera calibration and spectral smoothing have on sensing fecal contaminants on poultry carcasses. Spectral smoothing is a running average of spectra, in this work spectra is averaged over 20 nm. Four data pre-processing techniques were evaluated for the number of fecal and ingesta contaminants actually detected and the number of non-fecal or non-ingesta carcass features falsely identified as contaminants. The four pre-processing treatments were (1) no calibration, no spectral smoothing; (2) no calibration, spectral smoothing; (3) calibration, no spectral smoothing; and (4) calibration, spectral smoothing. Results on 64 corn fed birds indicated that the calibration with spectral smoothing was significantly better than the other three pre-processing methods for detecting fecal contaminants while minimizing false contaminant identifications. Now all future research with this type system will utilize calibration and spectral smoothing.
A hyperspectral imaging system was used to detect surface contaminants on 64 poultry carcasses fed a corn/soybean diet and subjected to a 53.3 degree C scald for 120 s. Hyperspectral data were analyzed with four pre-processing methods consisting of: uncalibrated data without spectral smoothing (uncalibrated raw); uncalibrated data with 20-nm spectral smoothing (uncalibrated smooth); percent reflectance calibrated data without spectral smoothing (calibrated raw); and percent reflectance calibrated data with 20-nm spectral smoothing (calibrated smooth). A common image-processing algorithm was then applied to each pre-processing method which included applying a background mask to the ratio a 565-nm image divided by a 517-nm image, and finally applying a fecal threshold. The results were then compared.
The calibrated smooth method was the best pre-processing method for contaminant detection. This is based on a high accuracy of 96.2% for predicting surface contaminants (although not significantly different [p<0.05] from two of the other methods) and significantly less (p<0.05) false positives (147). About 74% of the false positives identified by this method were from feather and boundary features of the carcass. Since only two wavelengths were used in the image-processing algorithm, it is suggested that another wavelength be examined that would identify feathers and thus, eliminate those false positives. The boundary features, which were identified as false positives, can be removed by slightly increasing the mask threshold which would prevent them from being included in the analysis.