Java library or strategy to tell if an image contains a signature

Question

I have an image file, and I need to determine if a specified area of this image contains a signature. Or to put it in end-user terms, "Has this document been signed?"

What I have done so far is to examine all the pixels contained in the area, to calculate an average "darkness", and compare that to a reference value. If the difference in darkness exceeds some threshold, then I consider it signed.

The problem with this (admittedly simplistic) approach is that because the pixels of the signature itself are such a small fraction of area, I have to use a very low darkness threshold, which results in a large number of false positives. I can't distinguish a real signature from stray markings, smudges, fax artifacts, etc.

To be clear...I'm not trying to match any specific signature or set of signatures. That is, I don't care who signed it, only whether it is signed.

Is anyone aware of a Java library that can do this, or of a better approach to this problem than what I am currently doing?

EDIT:

This is an example of the kinds of images I am working with. This document would be faxed to the recipient, signed and faxed back. It won't be this clean-looking by the time I need to look for a signature.

Answer 1

I do not know of any simple solutions. You can wrap over queXF or write something similar in Java. This paper talks about color code algorithm to recognize signatures.

Answer 2

This is what I believe can be done (although not a very good solution) but may still work. It would involve a bit of machine learning. I am assuming that your image does not contain hand written text and its just an image.

First thing to do would be to create a dataset of images which contain a signature and those which do not. The positive samples of the dataset should only contain signatures (you can learn a classifier for multiple aspect ratios) and negative samples should contain random images of the same aspect ratio/dimension. Now, you can compute some feature over these samples (HoG can be used as a feature, although I do not claim it is the best one to use for this application) and learn a SVM for each aspect ratio.

The next step would be to slide a detection window (of different aspect ratios) throughout the image and use the multiple SVMs you have learnt and check if any of them gives a positive response.

Although this approach may not work always, but should give a decent amount of accuracy. The more data you will use to learn, the better the results would get (and if you can come up with a good feature vector to represent a signature, it would help you case even further)