The same code run on the color advertisement database is run on the shape Chinese character database, but with a different interpretation of attribute. In the shape case, the attribute is the orientation of the ink along image curves.
The signature creation process is not performed directly on the bitmaps, but rather on the medial axis for the database characters since this makes it clearer exactly what are the image curves. The medial axis curves are divided into (relatively) small, equal length pieces. Each curve piece gives rise to a signel ((Aavg,Pavg),L) in the Orientation x Position image signature, where Aavg is the average orientation along the curve piece, Pavg is the average position along the curve piece, and L is the length of the piece.
The set of allowable transformations consists of scalings and translations. Since rotations are not allowed, the orientation of ink on the page guides or directs the search process. The system will not match ink at the same location on the page unless the orientations are relatively close. Two initial placements are used. The average query time over 15 different queries is Tavg=95.7 seconds, which is about 0.05 seconds per query-image comparison.
The results of a few queries are shown below. In each case, the query pattern is shown on the left, and the top 30 return images are shown on the right. We also show where SEDL believes the pattern occurs in a small sample of the returned images.
top | Title, Table of Contents, Introduction |
prev | Experiments with a Color Database of Product Advertisements |
The ideas and results contained in this document are part of my thesis, which will be published as a Stanford computer science technical report in June 1999.
S. Cohen. Finding Color and Shape Patterns in Images. Thesis Technical Report STAN-CS-TR-99-?. To be published June 1999.
Email comments to scohen@cs.stanford.edu.