GPU-Accelerating Denoising in 3D (GD3D) is an open-source implementation in CUDA of three commonly used image denoising methods: bilateral filtering, anisotropic diffusion, and non-local means. It was written to study the performance characteristics of these methods on 3D magnetic resonance images, as described in the paper:
Howison M and Bethel EW. GPU-accelerated denoising of 3D magnetic resonance images. (under review)
Download the latest
version from Bitbucket. The only prerequisite is the
CUDA Toolkit 5.0.35. You may
need to adjust the variable
CUDA_DIR in the
Makefile to point to your CUDA
installation. Also, you may need to change the line:
LIBS = -L$(CUDA_DIR)/lib -lcudart -lm
LIBS = -L$(CUDA_DIR)/lib64 -lcudart -lm
to build on a 64-bit Linux system (the included line should work on OS X).
Makefile will build six executables:
sweep/aniso sweep/bilat sweep/nlm tune/aniso tune/bilat tune/nlm
sweep executables perform a parameter sweep to determine optimal
smoothing parameters, while the
tune exceutables perform an auto-tuning
study to deterermine optimal memory blocking (see the paper for details).
GD3D is distributed under a modified BSD license. See LICENSE for full details.