EFFICIENT SAMPLING FOR THE EVALUATION PROTOCOL FOR 2-D RIGID REGISTRATION
Andreja Jarc, Peter Rogelj · 2009
In our research we aim to reduce the computation time spent on the evaluation protocol of a criterion function for rigid registration tasks. The basic evaluation protocol is performed on N uniformly distributed sampling lines in the K-dimensional transformation space. Similarity between two images is measured at each of the equidistantly placed points on each sampling line, which is a computationally intensive process. We hypothesize that the computational complexity can be reduced if attention is paid to the selection method of the sampling lines. In the research we compared four sampling methods which affect density and distribution of sampling lines: basic regular sampling, pseudo-random, Sobol quasi-random and Halton quasi-random sampling. We show that the use of Halton quasi-random generator yielded the most uniform distribution of sampling line. Thus, with proper sampling, the number of sampling lines could be systematically reduced in comparison to pseudo-randomly generated lines. This would result in shorter computation time spent on the protocol. The reduction of computation time is especially important when many criterion functions need to be evaluated (e.g. texture feature based image registration). The evaluation protocol was tested on a set of 11 2-D DRR (Digital Reconstructed Radiograph) and EPI (Electron Portal Image) image pairs. The tests have been conducted on intensity feature images as well as texture feature images extracted from the original intensity images. The paired Student's t-tests (p<0.05) indicated that results obtained from Halton quasi-random sampling were statistically significantly more consistent than the results based on regular or pseudo-random sampling.