Finding Compact Class Sets for Korean Font Image Classification

We address the problem of finding compact class sets for Korean font images taken under natural and noisy circumstances. Korean font images are prone to misclassification due to the similar, yet subtly different visual characteristics. The classification becomes even more confusing when the images are subject to various pixel-wise or affine translations, such as scaling and shear mapping. We argue that many font class divisions are inherently flawed in the sense that the fonts are divided in an overly-fine manner. To tackle this issue, we propose a system that discovers compact class sets, based on the confusion matrix of the initial classifier. We demonstrate that grouping existing classes into new ones increases the classification accuracy of Korean fonts, and also results in qualitatively intuitive new classes.