Skip to main content
ARS Home » Research » Publications at this Location » Publication #174962
Title: OPTIMIZATION OF FECAL DETECTION USING HYPERSPECTRAL IMAGING AND KERNEL DENSITY ESTIMATION

Author
item Yoon, Seung-Chul
item Lawrence, Kurt
item Park, Bosoon
item Windham, William

Submitted to: Transactions of the ASABE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/13/2007
Publication Date: 6/29/2007
Citation: Yoon, S.C., Lawrence, K.C., Park, B., Windham, W.R. 2007. Optimization of fecal detection using hyperspectral imaging and kernel density estimation. Transactions of the ASABE. 50(3): 1063-1071

Interpretive Summary: Fecal contaminants on poultry carcasses carry bacterial food-borne pathogens that are harmful to humans. A study was conducted to maximize the performance of a fully automatic fecal detection system using a real-time multispectral image system. The goal of the study was to find conditions on which the system minimizes the frequency of false fecal detections and also maximizes the detection accuracy. The study found that there was a theoretical performance limit to be achievable at a given detection accuracy with which a false detection rate could not be lowered below a certain level. A method was developed to predict a condition on which the detection accuracy could be maximized. A simple image ratio and the optimum threshold could discriminate feces and non-feces. The method to estimate the optimum threshold is important to improve the performance of a fecal detection system in poultry processing plants.

Technical Abstract: This paper demonstrated the development of a real-time multispectral image processing algorithm for surface fecal detection in poultry processing. Previous studies showed that a ratio of a 565-nm image to a 517-nm image could discriminate feces and non-feces. In this paper, the conditions for optimizing the image-ratio algorithm are reported. Applying a threshold after computing an image ratio is analogous to finding a linear decision boundary in statistical learning. This results in a trade-off between the detection accuracy and the false positives. To solve this problem, the duality of the Neyman-Pearson (NP) lemma was investigated, explaining the theoretical performance bounds achievable at the optimal conditions. In the NP framework, a lower bound of the detection accuracy was set while minimizing the false positives. An iterative numerical algorithm was also designed to solve the NP problem. For the design of the numerical solution, statistical density distributions of the fecal and non-fecal data were estimated by kernel density estimation, and characterized by edge models on a projection axis perpendicular to a linear decision boundary. Three necessary criteria were suggested for the selection of the optimum threshold in the image-ratio algorithm. Numerical simulations showed that the optimum threshold was 1.05. Optical trim filters were also designed maximizing the classification performance of a multispectral imaging system. Simulation results calculated with hyperspectral poultry images are provided.