Is there any source codes (whatever is the programming language) that allow to extract any features or structural primitives on handwritten words ? I'll use this features vectors as an input for k-means to cluster the words.

asked Feb 16 '12 at 16:45

shn's gravatar image

shn
462414759

The most basic one is the pixel grey density, which can be extracted by importing the image as an array. Is there a different type of feature you are after?

(Feb 16 '12 at 17:01) Robert Layton

@RobertLayton Only pixel grey density is clearly not sufficient to classify handwritten words. I'm interested in any features that can be helpful for handwritten words classification. I'm addressing words like those of this paper for example: http://www.cse.salford.ac.uk/prima/ICDAR2003/Papers/0163_533_pechwitz_m.pdf

(Feb 25 '12 at 05:47) shn

Well before doing any kind of "word clustering" (whatever that means) you need to segment the sequences of pixels as explained in the paper you mention. Why not implement the sliding windows over pixels feature extraction as explained the paper? It sounds very straight forward. The dynamic programming part to analyze the sequence less so though :)

(Feb 25 '12 at 10:03) ogrisel

I'm not sure it's straight forward, I didn't understood what are there features they extract exactly. What would be the feature vector of each data (image) ? If the images have not the same width, we obtain the same feature vector size for all images ?

(Feb 25 '12 at 11:18) shn

They are not feature vectors for each image, they are feature vectors for each possible vertical pixel position to find the optimal vertical splits to segment a single long image line into smaller images for each individual words.

I think the horizontal segmentation (to split individual text lines) is assumed to be solved "trivially" with classical image processing heuristics that do not involve machine learning in the preprocessing step. The Olena C++ library can do this part AFAIK.

(Feb 25 '12 at 12:11) ogrisel
Be the first one to answer this question!
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.