Exploiting Fisher Kernels in Decoding Severely Noisy Document Images

Decoding noisy document images is commonly needed in applications such as enterprise content management. Available OCR solutions are still not satisfactory especially on noisy images, and re-trainable systems require difficult and tedious training example preparation. Motivated by this challenging real application, we propose a novel solution that organically combines generative template models with discriminative classifiers via RBF Fisher kernel derived from a generative model. We show that the new approach is highly accurate in decoding noisy document images, making the system more generalizable to variations in font and degradation, and hence significantly reduces the burden in training example preparation. We also show that as it weights the pixel features by their relevancies, RBF Fisher kernel is more robust, and leads to smaller, faster models by dimensionality reduction.