In this master’s thesis, we investigate the use of neural networks for the
compression of dense volumes, which are essential for representing spatial
data in computer science and natural sciences. Due to their design, dense
volumes require significant storage space, a challenge that existing storage
solutions only partially address. We developed two neural compression ar-
chitectures: block-based and voxel-based, which combine the concepts of
autodecoders and coordinate neural networks. Both architectures offer ef-
ficient storage of dense volumes while enabling fast random access to their
values. For efficient implementation, storage, and data exchange, we devel-
oped a software library for the BVP format in C++ and extended it with
support for storing volumes in the proposed neural compression format and
the BC4 format. The VPT web visualization framework was upgraded with
support for decompressing neural-compressed volumes, enabling interactive
rendering directly in the browser. The results show that the proposed ap-
proach achieves high compression ratios with acceptable quality loss, with the
voxel-based architecture better preserving detailed structures. Comparison
with established BC4 and JPEG formats confirms the suitability of neural
compression as an effective alternative for storing volumetric data.
|
