Adaptable Similarity Search in Large Image Databases

Similarity has highly application dependent and even subjective characteristics. Similarity models therefore have to be adaptable to application specific requirements and individual user preferences. We focus on two aspects of adaptable similarity search: (1) Adaptable Similarity Models. Examples include pixelbased shape similarity as well as 2D and 3D shape histograms, applied to biomolecular and image databases. (2) Efficient Similarity Query Processing. Similarity models based on quadratic forms result in ellipsoid queries on highdimensional data spaces. We present algorithms to efficiently process ellipsoid queries on index structures, and improve the performance by introducing various approximation techniques that guarantee no false dismissals for both similarity range queries and k-nearest neighbor queries.