Recognition of Graffiti Tags

Paul Schwarz


Graffiti tagging, the inscription of an individual graffiti writer's unique mark in public space, is considered to be an antisocial behaviour and is a criminal offense. Due to the costs involved in the removal of graffiti tags it would be desirable to deter tagging. A searchable database of graffiti images could be an integral part of the policing system aimed at deterring graffiti vandalism. In order to index the database, an algorithm must be used that is capable of describing and comparing images. There are various approaches to object recognition and shape matching. The two methods that have been examined in this thesis are the Scale-Invariant Feature Transform (SIFT) and the Shape Context. These techniques were chosen based on their suitability to deal with the types of inconsistencies that are common in graffiti tag recognition, compared with the variances that occur in generic object recognition. Shape Context was found to be better suited as it was more robust than SIFT with regard to small shape distortions that are natural in graffiti writing. An empirical measure for the correctness of a searchable graffiti database was devised. Using this measure, it was found that retrieval results using Shape Context were considerably more accurate than retrieval results using SIFT. This thesis introduces the problem of recognising graffiti tags into the areas of object recognition and shape matching. Results are presented based on a collection of photographed graffiti tags.


Zip archive of graffiti images used in this work.