Building compact classifier for large character set recognition using discriminative feature extraction

In this paper, we propose an approach to building compact classifier for camera-based printed Japanese character recognition on mobile phones. We design feature vector prototypes using learning vector quantization (LVQ) for achieving high accuracy, while the complexity is lowered by linear dimensionality reduction. The discriminative feature extraction (DFE) strategy, which optimizes both subspace axes and classifier parameters, is shown to yield high classification accuracy even on low dimensional subspace. On a 120D sub-space, a 4,344-class classifier consumes only 613KB storage, and an accuracy of 99.41% was obtained on a test set.