YagpoOCRJan 13, 11:17Группа распознавания тибетского в Беркли приглашает нас к сотрудничеству
> Dear Friends,
> let me emphasize that we are entirely focused on the recognition of
> xylographs/blockprint. As of December 2007 we had our first end-to-end
> recognition system that took as input blockprint samples in TIFF
> have briefly summarized our results in a presentation available in the
> "Files" section of this newsgroup. The presentation is entitled:
> "TibetanOCRFlow-5slides-12-7-07."
> Briefly:
> 1) we continued to investigate a character-segmentation based
> approach; however, we were only able to achieve 92% segmentation
> accuracy. This obviously creates an upper bound on final recognition.
> 2) Our recognition algorithms are still pretty primitive and we only
> had ten pages of training data. As a result we only achieved about
> 80% accuracy.
> 3) Multiplying these limitations (and others) together our final
> recognition rate is only 70%.
>
> While not usable for anything other than coarse indexing, we still
> found these results encouraging as a first effort. We are currently
> investigating the use of generalized hidden-markov models to improve
> segmentation accuracy.
>
> One of my graduate students, Jike Chong, made another visit to Prof.
> Ding's group at Tsinghua University:
> http://www.csai.tsinghua.edu.cn/researchers/en_researchers_lab/dingxiaoqing.shtm
>
> While principally focused on the recognition of typeset Tibetan, we
> believe Prof. Ding's group continues to define the state-of-the-art
> in OCR for Tibetan. A few of Prof. Ding's papers are also available in
> the files section.
>
> You're welcome to join our google group at any time.
> If you are interested in cooperating with our efforts there are many
> activities that can be explored independently.
>
> Kind regards,
> Kurt
> http://www.eecs.berkeley.edu/~keutzer/
>
>
>