Thursday, January 10, 2013

Improving Dictionary Learning: Multiple Dictionary Updates and Coefficient Reuse - implementation -


Leslie just sent me the following:

Dear Igor,

Since sparse representations is a related topic you discuss on Nuit Blanche, I would like to mention a recent joint paper with Michael Elad on improving dictionary learning that has recently been published in IEEE Signal Processing Letters. The paper is titled "Improving Dictionary Learning: Multiple Dictionary Updates and Coefficient Reuse" and can be download at this link: http://www.cs.technion.ac.il/~elad/publications/journals/2012/IEEE-SPL-DL-2012.pdf 
Perhaps even more interesting is that we have made available the software for the methods described in the paper and information on this software can be found at http://www.cs.technion.ac.il/~elad/software/  
Thanks for all your efforts with the blog and related study groups (LinkedIn, Google+, etc.).
Best regards,
Leslie
Thanks Leslie.


In this paper we propose two improvements of the MOD and K-SVD dictionary learning algorithms, by modifying the two main parts of these algorithms – the dictionary update and the sparse coding stages. Our first contribution is a different dictionary-update stage that aims at finding both the dictionary and the representations while keeping the supports intact. The second contribution suggests to leverage the known representations from the previous sparse-coding in the quest for the updated representations. We demonstrate these two ideas in practice and show how they lead to faster training and better quality outcome.


2 comments:

Anonymous said...

hi,
thanks for sharing the code !!!!
but i could not find it the your specified web page.
Can you check it again please?

regards

Igor said...

At the very least you could go on Miki Elad's page: http://www.cs.technion.ac.il/~elad/software/ as mentioned in the blog entry.

Igor.

Printfriendly