A new approach for detection of stamps in color document
images was presented in this paper. It involves conversion
into Y CbCr color space and extraction of chromatic pixels.
Ink color clusters in Y CbCr have elongated shapes and
this property is exploited for image segmentation by color
clustering. Candidate solutions are obtained and classified
based on several geometrical and color-related features.
Stamps of any shape and arbitrary colors except for black
can be detected by the algorithm. Nevertheless, the method
is extensible for detection of black stamps too. A more
accurate segmentation algorithm would have to be applied
to the achromatic part of the image, the same geometrical
features could be used and new features e.g. concerning the
printing differences would have to be added.
A thorough analysis of previous work has been done and
we claim the here presented approach to stamp detection to
be the most generic one so far.
A new data set of 400 documents with stamps of different
shapes and colors has been collected and made public. The
method was evaluated on it with recall as well as precision
of 83%. Evaluation on documents with low resolution corresponding to 200 dpi scans lead to the same results.
As a part of ongoing work, preliminary experiments suggest that the method can also distinguish between authentic
stamps and some photocopies