Recognition of Ultra Low Resolution Word Images Using HMMs.
Farshideh Einsele, Rolf Ingold, Jean Hennebert · Computer Recognition Systems · 2008
We report in this paper on significant improvements that we have been including in our HMM-based system for recognition of ultra low resolution, antialiased text with small point sizes such as those frequently found in web images. First we are proposing a fully automatic training procedure where no a priori knowledge of font metrics is needed. This means that our system can be potentially built to recognize any font. Second, the system’s performance can be boosted by using mixtures of Gaussians to model the probability density functions of the HMMs. Third, we show that these improvements allow the system to handle large vocabulary size up to 60’000 wordswith few degradation of the accuracy. We report on these results for 2 different font families, namely a serif and a sans serif font. We also report on different HMMs topologies and conclude on the benefits of using minimum duration to model the characters composing the words.