A Swig-Python Wrapper for Video Non-Local Bayesian Denoising (C++ code originally from Pablo Arias)
$ git clone https://github.com/gauenk/svnlb/
$ cd svnlb
$ python -m pip install -r requirements.txt --user
$ ./install.sh
We expect the images to be shaped (nframes,channels,height,width) with
pixel values in range [0,...,255.]. The color channels are ordered RGB. Common examples of noise levels are 10, 20 and 50. See scripts/example.py for more details.
import svnlb
import numpy as np
# -- use enough threads --
svnlb.check_omp_num_threads()
# -- get data --
clean = 255.*np.random.rand(5,3,64,64)
# (nframes,channels,height,width)
# -- add noise --
std = 20.
noisy = np.random.normal(clean,scale=std)
# -- TV-L1 Optical Flow --
fflow,bflow = svnlb.swig.runPyFlow(noisy,std)
# -- Video Non-Local Bayes --
result = svnlb.swig.runPyVnlb(noisy,std,{'fflow':fflow,'bflow':bflow})
denoised = result['denoised']
# -- compute denoising quality --
psnrs = svnlb.utils.compute_psnrs(clean,denoised)
print("PSNRs:")
print(psnrs)The outputs from the Python Wrapper and the C++ Code are exactly equal. To demonstrate this claim, we provide the scripts/compare_cpp.py script. We have two examples of the C++ Code output ready for download using the respective scripts/download_davis*.sh files. To run the data downloading scripts, type:
$ ./scripts/download_davis_64x64.sh
To run the comparison, type:
$ export OMP_NUM_THREADS=4
$ python scripts/compare_cpp.py
The script prints the below table. Each element of the table is the sum of the absolute relative error between the outputs from the Python Wrapper and C++ Code.
| noisyForFlow | noisyForVnlb | fflow | bflow | basic | denoised | |
|---|---|---|---|---|---|---|
| Total Error (cv2) | 0.000505755 | 0 | 504.308 | 21.643 | 0 | 0 |
| Total Error (cpp) | 0 | 0 | 0 | 0 | 0 | 0 |
Details can be found in docs/COMPARE.md
The code depends on the following packages:
- CBLAS, LAPACKE: operations with matrices
- OpenMP: parallelization [optional, but recommended]
- libpng, libtiff and libjpeg: image i/o
NOTE: By default, the code is compiled with OpenMP multithreaded
parallelization enabled (if your system supports it). Use the
OMP_NUM_THREADS enviroment variable to control the number of threads
used.
This code provides is a Python wrapper over an implementation of the video denoising method (VNLB-H) described in:
Please cite the publication if you use results obtained with this code in your research.
- Original Code linked here
- C++ (and Primary) Author: Pablo Arias [email protected]
- For computing the optical flow, it includes the IPOL implementation of the TV-L1 optical flow method of Zack and Pock and Bischof.
- For image I/O, we use Enric Meinhardt's iio.
- For SWIG-Python, Kent Gauen wrote this wrapper [email protected]
Licensed under the GNU Affero General Public License v3.0, see LICENSE.