Speeding up local patch dissimilarity

Abstract

There are many patch-based techniques used in image processing, but most of them are heavy to compute with current machines. A dissimilarity measure for images based on patches, inspired from rank distance, called Local Patch Dissimilarity (LPD), was recently introduced. It has very promising results in optical character recognition, but, as other patch-based methods, it is computationally heavy. This work aims at showing that LPD can be improved in terms of efficiency. Several ways of optimizing the LPD algorithm are presented, such as using a hash table to store precomputed patch distances or skipping the comparison of overlapping patches. Another way to avoid the problem of the higher computational time on large sets of images is to turn to local learning methods. Several experiments are conducted on two datasets using both standard machine learning methods and local learning methods. All methods are based on LPD. The obtained results come to support the fact that LPD is a very good dissimilarity measure for images. In this paper, LPD is also used with success for classifying images other than handwritten digits. © 2013 Springer-Verlag.