Avi Bleiweiss
Optical character recognition, histogram of oriented gradients, bootstrapping, artificial warping, support vector machines
We present a novel, high performance character recognition system for handwritten Hebrew scripts. Specifically, we explore a technique that bootstraps from an extremely limited hand-made seed of examples, to orderly sample a much amplified synthetic set, at training runtime. This flexible formulation, effectively trades off learning performance vs. memory footprint and scales down to fit the resource limited, small form factor compute devices. We exploit a compressed format of histogram of gradients (HOG) features, trained on a one-against-all SVM classifier, and our evaluation shows a markedly high accuracy rate when compared to other OCR systems.
Important Links:
Go Back
