Skip to main content
ARS Home » Northeast Area » Kearneysville, West Virginia » Appalachian Fruit Research Laboratory » Innovative Fruit Production, Improvement, and Protection » Research » Publications at this Location » Publication #161366


item Peterson, Donald
item Tabb, Amy

Submitted to: Transactions of the ASAE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/13/2005
Publication Date: 12/1/2007
Citation: Bennedsen, B.S., Peterson, D.L., Tabb, A. 2007. Identifying apple surface defects using principal components analysis and artificial neural networks. Transactions of the ASAE. Vol. 50(6), 2007 (ASABE) American Society of Agricultural and Biological Engineers ISSN 001-2351.

Interpretive Summary: The work was based on a sorting system for apples, developed over the past years, in collaboration between Cornell University and the Appalachian Fruit Research Station. The system included an image acquisition system, which captured images at two near infrared wavelengths, 740 nm and 950 nm respectively. These wavebands had been found to be particularly suited for detection of blemishes from diseases (740 nm) and bruises (950 nm). Principal components and neural networks were used in an attempt to detect the defect in the near infrared images. Training sets of up to 18 apple images were constructed and used to train neural networks. The networks were tested on a standard set of 20 apple images with different defects. In all cases, the number of defects detected by the network as well as the defective area was recorded. The effect of subjecting the images to different pre-processing before extracting principal components and training the network was examined. The background was eliminated and the apple surface stretched by an image processing routine to fit the frame. Further, segments of the frames were extracted and combined to form images consisting of six frames and covering exactly 360 degrees of the apple surface. A Wiener filter was employed in order to reduce local variations in grey levels. The best result obtained was detection of up to 79% of the defective surface of the apples in the test set. While most of the larger defects (91%) and bruises were detected (92%), small dark spots and faint marks were not adequately caught by the neural networks. The conclusion is that the method in itself is not sufficient for practical implementation.

Technical Abstract: Artificial neural networks and principal components were used to detect surface defects on apples in near infrared images. Neural networks were trained and tested on sets of principal components derived from columns of pixels from images of apples acquired at two wavelengths; 740 nm and 950 nm. Different ways of preprocessing images prior to training the networks were attempted. Results could be improved by removing the background and applying a Wiener filter to the images. Overall, the best result obtained was 79% of the defects detected in a test set consisting of 185 defects.