On Wednesday, June 6, 2007 2:19:31 AM UTC+9, Qaaddi wrote:
> Ganesan Ramachandran wrote:
> >
> >
> > you can use kmeans to come up with state prototypes. usually the
> > transition
> > matrix is either randomly initialized or uniformly initialized.
> >
> >
> > "Qaaddi" <bassqqad@yahoo.com> wrote in message
> > news:ef58fe2.1@webcrossing.raydaftYaTP...
> >> I'm using the HMM toolbox for a data. As you know, MATLAB has a
> >> "hmmtrain" function that gives you estimates of the hmm
> > parameters.
> >> but, to use this code, you need to have a very good initial
> guess
> > of
> >> the parameters. my question is what if the only thing you have
> is
> > the
> >> data string itself(the observations) and nothing more?
> >> then how can
> >> you come up with a good initial guess for the transmission and
> >> emission matrices?
> >
> *****************************
> Dear Mr.Ramachandran
> Thanks for reply.
> Actually i have a codebook of quantized vectors as my observations.I
> want to recognize 15 patterns.
> each pattern is a series of a number of codewrods.
> I want to train 15 models to recognize 15 patterns and all what i
> have is the sequence of codewords(observations).
>
> The function [hmmestimate] needs the state sequence which i dont have
> , and the function [hmmtrain ]needs a guess of initial Transitions
> and Emissions matrices which also i dont have
>
> Sorry, i didnt understand how kmeans will help me?
> plz, try to explain alittle .
> Thanks alot
Hi Qaaddi how do you quantized vector ?
I am working in character recognition. below is my properties of input image
type : binary image
width : 400
height : 100
then i divided each image into 10 horizontal blocks. After that i calculate the number of black pixels in each block.
Suppose this is my feature vector.
feature=[1 2 3 4 5 6 7 8 9 10];
how do i train using matlab hmmtrain command and how to define emission matrix.
Thanks
