We present a hybrid CPU-GPU approach for the Fourier-based stitching of optical microscopy images. This system achieves sub-minute stitching rates with large grids; it stitches a grid of 59x42 tiles in 26 seconds on a two-CPU (8 physical cores) & two-GPU machine. This is a speedup factor of more than 24x; the optimized sequential implementation takes more than 10 minutes to perform the same task. The system scales to take advantage of additional CPU cores or GPU cards. For the sake of comparison, ImageJ/Fiji, which uses a similar algorithm, exceeds 3.6 hours on the same workload.
Proceedings Title: GPU Technology Conference 2013
Conference Dates: March 18-21, 2013
Conference Location: San Jose, CA
Pub Type: Conferences
Image Stitching, Multi-core computing, GPGPU