|
|
|
|
Mining Image Content by Aligning Entropies with an Exemplar
Clark F. Olson In Multimedia Data Mining and Knowledge Discovery, eds. V. A. Petrushin and L. Khan, pages 325-339, 2007. In order to efficiently answer queries on image databases at run-time, content must be mined from the images offline. We describe a technique for locating objects in the library for which we have an exemplar to compare against. We match the images against the exemplar by comparing the local entropies in the images at corresponding positions. This representation is invariant to many imaging phenomena that can cause appearance-based techniques to fail. It can succeed even when the images are captured with different sensors (for example, CCD versus FLIR). We employ a search strategy that combines sampling in the space of exemplar positions, the Fast Fourier Transform (FFT) for efficiently evaluating object translations, and iterative optimization for pose refinement. Experiments indicate that the sampling can be somewhat coarse. The techniques are applied to matching exemplars with real images. We describe strategies for scaling this approach to multimedia databases and conclude with a discussion of future work that would be beneficial for mining content from images and video data. |
