GPU implementation of volume reconstruction and object detection in Digital Holographic Microscopy

Abstract

Using Digital Holographic Microscopy (DHM) we can gather information from a whole volume and thus we can avoid the small depth of field constraint of the conventional microscopes. This way a volume inspection system can be constructed, which is capable to find, segment, collect, and later classify those objects that flow through an inspection chamber. Digital hologram reconstruction and processing, however, require considerable computational resources. We are developing volume reconstruction and object detection algorithms that can speed up considerably by parallel hardware implementation. Therefore, we put these tasks into operation on a GPU. As data transfer of the reconstructed planes would slow down the algorithm, all the reconstruction, object detection processes are to be completed on the parallel hardware, while fine tuning of object reconstruction and classification will be done on a CPU later. The actual speed up of the GPU implemented algorithm comparing to its conventional CPU realization depends on the applied hardware devices. So far we reached a 10 times acceleration value.