Skip to main content
ARS Home » Pacific West Area » Albany, California » Western Regional Research Center » Foodborne Toxin Detection and Prevention Research » Research » Publications at this Location » Publication #206249
Title: An Automatic Algorithm for Detection of Inclusions in X-ray Images of Agricultural Products

Author
item Haff, Ronald - Ron
item Pearson, Thomas

Submitted to: Electronic Publication
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/26/2007
Publication Date: 7/18/2007
Citation: Haff, R.P., Pearson, T.C. 2007. An Automatic Algorithm for detection of Infestations in X-ray Images of Agricultural Products. Sensing and Instrumentation for Food Quality and Safety. 1(3):143-150

Interpretive Summary: A computer program was developed and tested for detection of certain defects or contaminants in x-ray images of food products The program was tested on x-ray images of wheat kernels infested with larvae of the granary weevil and the results were compared to those obtained by human subjects evaluating digitized x-ray film images (14.4% overall error vs. 15.6% for human subjects). The program was also tested on x-ray images of olives infested with the Olive Fly, yielding a total error of 12% for large infestations and over 50% for the smallest infestations with false positive (good product classified as bad) results below 10%. Certain training strategies for the computer program were derived and tested.

Technical Abstract: An automatic recognition algorithm was developed and tested for detection of certain defects or contaminants in x-ray images of agricultural commodities. Testing of the algorithm on x-ray images of wheat kernels infested with larvae of the granary weevil yielded comparable results to those obtained by human subjects evaluating digitized x-ray film images (14.4% overall error vs. 15.6% for human subjects). Further testing on x-ray images of olives infested with the Olive Fly yielded a total error of 12% for large infestations and over 50% for the smallest infestations with false positive results below 10%. Testing of alternate training strategies showed that for this type of algorithm, which uses a form of discriminant analysis with a generally “fuzzy” decision boundary, best results are obtained training with samples that map far away from the boundary, then applying the derived decision function to all samples to be classified.