filtering text and non text region based on svm classification

My work is on text detection. I used connected component analysis. Now i want to filter the nontext and nontext regions.I used to avoid the region based on area.But it does not work on all images. So I think svm is suited for this by takingg the probability that the character present in the set,the set contains all the characters like A..Z.a..z,0...9. I am new on MATLAB. I saw the example given in the matlab But I didint get it. How the training and testing samples are formed in this case. What are steps? Features are extracted on training or testing samples? please help me