Pattern search in a shoe sole image database using eigenpatterns

This paper suggests a method for automating the process of checking for a sufficiently matching sole pattern in a given collection of sole patterns. It may take up to weeks to manually do the same, in the industry. We describe the method of finding the best match from a given set using a procedure, the basis of which is the Karhunen-Loueve transform (KLT), involving a projection onto a subspace spanned by the eigenpatterns corresponding to the largest few eigenvalues. The patterns are first scanned, stored in picture file format, and cleaned up so that they consist of a black sole and a white background, to minimize unnecessary information content. Their centers of masses are aligned and principal moment of inertia axes are arranged vertically to maximize correlation between these pictures. The eigenpatterns are found for this collection of pictures. To find the best match for the new sole pattern, its projection onto the lesser-dimensional subspace spanned by the eigenpatterns is found. The projected pattern can be decomposed into a weighted sum of the eigenpatterns. The best match is found using a minimum distance criterion in this space. This method is suitable for any highly correlated image database. © 2003 Elsevier Science Ltd. All rights reserved.