Speeding up IDM Without Degradation of Retrieval Quality

Michael Springmann, Heiko Schuldt
In Proceedings
Appears in
Advances in Multilingual and Multimodal Information Retrieval:Proceedings of the 8th Workshop of the
Budapest, Hungary
Springer Berlin / Heidelberg
The Image Distortion Model (IDM) has previously shown good retrieval quality. However, one of the limitations that may limit its use in a wider range of applications is computational complexity. In this paper, we present an approach that applies several optimizations to decrease the retrieval time of IDM without degrading the quality of query results. We were able to perform the IDM in less than 1.5 seconds per query on an 8-way server and 16 seconds on a standard Pentium 4. In particular, the early termination strategy we applied contributed a speedup of up to 4.9. We also extended the possible displacements to an area of 7 ×7 pixels with a local context of up to the same size. The results submitted to the medical automatic annotation task of ImageCLEF’2007 were ranked in the upper third. Most importantly, the proposed techniques are not limited to IDM but can also be applied to other expensive distance measures.