3
$\begingroup$

Is there any interesting problem in the area of "Document Image Analysis and Retrieval" which by nature needs an online/incremental clustering process ? The problem may be in the context of "Logical Structure Analysis", or "Document Layout Analysis" to identify regions of interest in a scanned page, or any other related topics. What matters is that the considered problem naturally needs an online/incremental clustering. Do you have any ideas or suggestions about such problems ?

Note: the considered document images are actually a scanned administrative documents

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

If you use the term "clustering" in the sense of "near duplicate detection", the online updating of IR duplicate indexes is an obvious candidate.

Think of Google image search that wants to merge duplicate images as they are spidered and coming into the index (instead of bulk-rebuilding the index, as everybody used to do).

If you use a broader term of clustering, such as "related but not identical objects" it is a bit harder to find something. Try to think of actual data sources that produce a stream of images that is not continuous video...

Improve this answer
$\endgroup$
5
  • $\begingroup$ Suppose you are a company and you have a big continuous flow of administrative documents (scanned) that comes every day. One problem could be to automatize the segmentation of this documents on the fly (identifying physical and logical regions of interest), so by clustering this flow of documents online, one solution of segmentation of one document could be directly applied to similar documents. I trying to define a subject concerning problems like that. $\endgroup$
    – shn
    Commented Sep 11, 2012 at 20:26
  • $\begingroup$ Clustering is probably less efficient and less effective than the actual segmentation or classification. Because in this setting, you do have training data; and then supervised methods always outperform unsupervised methods. And most likely, the key idea is to use OCR to turn it into text and work with that. Because the majority of documents will be printed, and then OCR works quite good, at which point you are in text and no longer in images. $\endgroup$
    – Has QUIT--Anony-Mousse
    Commented Sep 11, 2012 at 20:34
  • $\begingroup$ We can not use supervised methods. We don't have a previous fixed training dataset. We have as input a stream of documents that comes continuously and we also can receive new documents that have never been seen before, so an online clustering -I think- is what we should do here. Now when we have a new document that should be assigned to a given cluster representative, we can apply the solution of segmentation of this representative (which is probably the same for all documents assigned to this cluster) $\endgroup$
    – shn
    Commented Sep 11, 2012 at 20:42
  • 1
    $\begingroup$ Well, do you need to do it on an image processing level? Why not OCR? And IMHO it pays off more to obtain training data than to fully bet on unsupervised methods. They may produce clusters that make mathematical sense, but are not useful, after all. One cannot assume that all clusters found are meaningful/helpful to the users. $\endgroup$
    – Has QUIT--Anony-Mousse
    Commented Sep 12, 2012 at 6:02
  • $\begingroup$ I can use OCR and represent each new document as a vector (bag of words), then the online clustering is applied on this vectors. For each existing representative a "segmentation solution" is associated which tell as how to segment documents that are similar to this representative. This is feasible, isn't it ? I can not have an prior fixed training data, since the stream of coming documents can evolve any-time $\endgroup$
    – shn
    Commented Sep 12, 2012 at 11:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.