FILTER DESIGN FOR OPTIMAL FEATURE EXTRACTION FROM X-RAY IMAGES

Abstract

X-ray linescanning is beginning to be established as an alternative tool for non-destructive produce inspectionwhen internal defects are expected. Image analysis based techniques are being increasingly used for quality evaluationand classification of fresh fruits and vegetables. However, noise removal and feature extraction techniques suitable forparticular applications need to be identified before a classification system can be developed. This study focuses on thedesign of prefiltering steps which will lead to optimal defect detection. Autocorrelation and variogram analyses wereperformed on water CT scans, water linescans, and plexiglass linescan images for purposes of quantifying instrumentrelated noise. Little correlation was observed in water CT and plexiglass linescan images, which implied that a Gaussianfilter may be appropriate for noise removal from linescan images of fruits prior to removal of edge features associatedwith defects such as bruises. A 5 5 Gaussian kernel gave significantly (P 0.05) better discrimination based onexperimental results for enhancing bruise features in apples and disease features in onions. This result confirmed a rule ofthumb in terms of filter design based on image statistics. In instances where morphological operators were used,prefiltering masks were not found to be significant at the 5% level. There was a strong correlation between the imagesignal and pixel location in linescans of water with a free surface. This correlation is believed to be caused by waves dueto motion, suggesting a need for frequency domain filtering in the case of water transport.