Konferenční příspěvek

Horáček O., Bican J., Kamenický J., Flusser J.

Computer Recognition Systems

2

Springer

Berlin

2007

978-3-540-75174-8

Ing. Ondřej Horáček, Ph.D. (ÚTIA - Oddělení zpracování obrazové informace)

Image identification, Image database, Decision tree, Invaria

Image retrieval deals with a problem of finding similar pictures in image database. Our task is to find originals of modified images, typically stolen and republished on the web. Our problem is specific in terms of the database size (millions of photos), demanded speed of the search (seconds), and unknown image modifications (loss of quality, radiometric degradation, crop, etc.). Proposed method works in the following tree steps: 1. Image preprocess -- normalization for robustness to the modifications. 2. Retrieval of candidates from the database index -- stochastic decision in each vertex of the index tree is used to find the most relevant candidates. 3. Verification of the candidates -- modified phase correlation is used. The method was implemented in practice with very good results. Based on wide experiments, it was shown that the success rate of the search depends on the level of image modification.