Labeled images verification using Gaussian mixture models

Abstract

We are proposing in this paper an automated system to verify that images are correctly associated to labels. The novelty of the system is in the use of Gaussian Mixture Models (GMMs) as statistical modeling scheme as well as in several improvements introduced specifically for the verification task. Our approach is evaluated using the Caltech 101 database. Starting from an initial baseline system providing an equal error rate of 27.4%, we show that the rate of errors can be reduced down to 13% by introducing several optimizations of the system. The advantage of the approach lies in the fact that basically any object can be generically and blindly modeled with limited supervision. A potential target application could be a post-filtering of images returned by search engines to prune out or reorder less relevant images.

GMM-based handwriting style identification system for historical documents

“…The handwriting documents used in this work are from the 17 th century and available in this link 10 . Figure 1 illustrates one sample page of each of these handwriting styles.…”

“…Gaussian Mixture Models (GMMs) were used in many domains such as font recognition [8], [9], label image verification [10], baseline estimation [11], writer identification [12], etc. In this work GMMs are used to model the handwriting style of each historical book(s).…”