logic of recognition

[apinjapan]

dear all,
as this project seems to be open source, i was hoping someone could
answer a few basic questions. I'm not a real programmer, and don't
know anything about python. I was just hoping someone could lay out
the basic logical steps to achieve the recognition. What gets compared
to what, in what order. something like that. thanks.

[Mathieu Blondel]

Hello,

Apologies for the very late answer. The basic idea used in Tegaki is
to rely on character examples. Currently, Tegaki includes 2
recognizers, Zinnia and Wagomu. Zinnia is an example of model-based
approach: the character examples are compiled into a model (this
process is often called "learning" or "training" in the machine
learning community, or just "fitting" in the statistics community) and
the recognizer uses the model to make predictions. On the other hand,
Wagomu is an example of instance-based approach: the recognizer uses
character examples directly; it measures the similarity between the
input character and the character examples.

Zinnia is based on Support Vector Machines (SVM) while Wagomu is based
on Dynamic Time Warping (DTW). In Zinnia, each character is
represented by a point in a n-dimensional space. Each dimension
corresponds to a feature describing the character. The goal of the
model fitting problem is to find the optimal weight of each feature
(the ones which maximizes classification accuracy). At recognition
time, the score of each character is computed by taking the weighted
sum of the features and the character with the greatest score is
selected. DTW is an example of elastic metric: unlike the euclidean
distance, even if two characters or strokes have different lengths, it
can output a distance value between these two. Wagomu compares the
input with each candidate character and select the one with smallest
distance.

HTH,
Mathieu

[ehrior]

Are there some papers or books that newbie should read to understand
the logic?

Yesterday I started to use Mendeley program, a software for managing
and sharing research papers and there is an option to share papers, so
we can try to use it for papers that newbies should read.

Thanks.
David

[Mathieu Blondel]

On Aug 19, 7:33 pm, ehrior <lucaste...@gmail.com> wrote:
> Are there some papers or books that newbie should read to understand
> the logic?

Here's a Google search which should retrieve a few relevant papers:

survey (kanji OR japanese OR chinese) online handwriting recognition

The most popular approaches are based on Hidden Markov Models (HMMs).

Mathieu