Re: zinnia accuracy
Mathieu Blondel
Shinkanji iphone app) sent to the author of Zinnia and me. This may
explain why Zinnia performs so poorly at recognizing individual
strokes.
On Fri, Sep 18, 2009 at 4:10 PM, Benoit Cerrina <benoit...@gmail.com> wrote:
> Hello Matthieu, Taku,
> I am trying to improve the quality of the zinnia recognition and the cpu
> consumption, to do this I would like to study what are the most important
> "features" for Zinnia.
> As a reminder, this is my analysis of how zinnia proceeds to extract a
> vector of features describing a character.
> 1) for each stroke zinnia is extracting the meaningfull stroke pairs.
> Features::getVertex starts with the 2 end points of the stroke to analyze
> (A,B), this 2 end points make up the first meaningfull pair for the stroke.
> Then the intermediary point C which is at the greatest distance from the
> segment is estimated using a distance approximation. If this estimation is
> greater than a predefined threshold the segments (A,C) (C,B) are considered
> to be meaningfull.
> 2) for each meaningfull stroke pair the following feature vector is created
> (in Features::makeBasicFeature)
> 1: length of the segment
> 2: angle of the segment compared to the x axis
> 3,4,5,6: cartesian coordinates of the start and end points of the segment
> (with the origin placed at the center of the kanji and the width/height of
> the kanji being equal to 1)
> 7,8,9,10: polar coordinates of the start end end points (7, 8 are the angles
> 9, 10 the distance)
> 11,12: length of the projection of the segment on the x and y axis
> The weigths assigned to each of those elements are:
> 1,3,4,5,6,9,10 *10
> 2,7,8 *1
> 11,12 *5
> 3) for each segments between strokes (starts at the end of one and finish at
> the start of the next) the same feature vector is extracted
> I am trying to figure out if I can tweak the different weights and remove
> some features to improve accuracy and speed.
> Would you have any suggestions
> Best regards
> Benoit Cerrina