Large Histogram Computation for Normalized Mutual Information on GPU

This poster presents two GPU implementations of normalized mutual information (NMI) for a two-step image registration used in version 1.0 of the Photogrammetric Registration of Imagery from Manned and Unmanned Systems (PRIMUS) pipeline [1]. This fully-automated, sensor-agnostic application enables the precise geolocation of aerial and orbital imagery, a fundamental building block for accurate geographic applications and data analysis. In PRIMUS, NMI is used for coarse and refined alignment of multi-spectral imagery, referred to as global localization and registration, respectively. However, different algorithms have been implemented in CUDA to fit the different characteristics of each alignment. Although NMI is commonly used to register images from different modalities [2], GPU implementations usually focus on single coefficient computation leveraging entropy approximation calculus or generating histograms with a limited number of bins [3][4][5]. However, for our use case, we need to generate the full joint-histogram of 65,536 bins to mitigate the loss of accuracy due to compression that occurs when reducing the image-intensity range from 16-bit to 8bit, needed to leverage most of the functions in OpenCV library. It is also important to note that our use case involves a mask that indicates valid pixels that are considered for the NMI coefficient computation this adds to the computational complexity and uniqueness of our solution but can significantly improve its overall accuracy. Our preliminary approaches target the NVIDIA Tesla K80, not only to leverage the global memory available on the GPU but also its queue manager, which lends themselves well to our specific problem.