Improving efficiency and effectiveness of the image distortion model

The image distortion model (IDM) is a deformation model that is used for computing the (dis-)similarity between images. Therefore it evaluates displacements of individual pixels between two images within a so-called warp range and also takes into account the surrounding pixels (local context). It can be used with a kNN classifier and has shown good retrieval quality in handwritten character recognition as well as in past runs of the medical automatic annotation task of ImageCLEF workshops. However, one of its limitations is computational complexity and the resulting long query response times, that may limit its use for a wider range of applications and for modifications to further improve retrieval quality. In particular an enlarged local context and warp range are candidates for such improvements, but would even further increase computational complexity. In our approach, we therefore apply several optimizations that reduce the retrieval time without degrading the result quality. First, we use an early termination strategy for the individual distance computations which contribute a speedup of a factor of 4.3-4.9. Second, we make efficient use of multithreading. With these extensions, we are able to perform the IDM in less than 1.5s per query on an 8-way server and 16s on a standard Pentium 4 PC without any degradation of retrieval quality compared to the non-optimized version. We extend the possible displacements to an area of 7x7 pixels, using a local context of either 5x5 or 7x7 pixels. The results of the extended IDM have been submitted to the medical automatic annotation task of ImageCLEF 2007 and were ranked in the upper third. More importantly, the used techniques for reducing the execution time are not limited strictly to IDM but are also applicable to other expensive distance measures.