Abstract:

Image registration is the process of finding pixel correspondences between images that were taken at different points in time or from different viewpoints. In particular, we are interested in registering exposures captured with varying shutter speed settings for the creation of high dynamic range (HDR) images. For an existing histogram-based image registration technique, we discuss the suitability for parallelization. Modifications to the technique that allow us to run parts of the algorithm in parallel are presented. The parallel parts can then be processed by a graphics processing unit (GPU) for improved performance. Our results show that the most time consuming component of the algorithm can be sped up on an Nvidia GPU by a factor of up to 19 over the original sequential version. The average speed-up is 5.9:1.

Paper available at IEEE.