Friday, March 13, 2009
Duhu (Dumb Human!)
A script that should identify a segmented letter with a certain confidence. If its confidence is above a certain value, then it should be right!
Now using this as a feedback mechanism, we try to do the segmentation and identification of just one letter. Now we decrease the segmented length and then continue detection for another letter...
Main goal is to go for a feedback based mechanism...
SVM, KNN or Neural: All the 3 seem to be similar in terms of performance. SVM and KNN cannot perform rotations unless they are trained on samples with rotation. Convolutional Neural Networks can do the rotation.
FEATURE EXTRACTION:
Calculate features using vectors. i.e. "T" means one vector to the right, the one down, and one to the right again. Store the starting position (x, y, z) and the vector (i, j, k). You can have about 5 such vectors for each character. We can scale down all vectors of a letter to some value (maybe constant height) and have an extra feature called scale for each letter. This way we can recognize letters of varying sizes. For, letters where all 5 are not utilized we can have 0s.
And, if we can rotate these vectors we will be able to make it resistant to rotation!! But how exactly does it cope with distortion and waving?
Slowing down... because of pe_dump. Will do this in my free time
Now using this as a feedback mechanism, we try to do the segmentation and identification of just one letter. Now we decrease the segmented length and then continue detection for another letter...
Main goal is to go for a feedback based mechanism...
SVM, KNN or Neural: All the 3 seem to be similar in terms of performance. SVM and KNN cannot perform rotations unless they are trained on samples with rotation. Convolutional Neural Networks can do the rotation.
FEATURE EXTRACTION:
Calculate features using vectors. i.e. "T" means one vector to the right, the one down, and one to the right again. Store the starting position (x, y, z) and the vector (i, j, k). You can have about 5 such vectors for each character. We can scale down all vectors of a letter to some value (maybe constant height) and have an extra feature called scale for each letter. This way we can recognize letters of varying sizes. For, letters where all 5 are not utilized we can have 0s.
And, if we can rotate these vectors we will be able to make it resistant to rotation!! But how exactly does it cope with distortion and waving?
Slowing down... because of pe_dump. Will do this in my free time
Fox For All | Blogger Templates by Gecko &
Fly.
No part of the content or the blog may be reproduced without permission.
