Blog
Post
|
Report Bug on Numpy Version
|
Report Bug on PyOpenCL Version
This repository is the Python implementation of the paper: Image Smoothing via L0 Gradient Minimization
| Flower |
|---|
 |
| Rock |
|---|
 |
pip install L0-Smoothing# Clone the repository
git clone https://github.com/TsXor/L0-Smoothing.git
# build package
# build.bat is batch script for Windows cmd, and will not work on linux.
# build.bat consist of mostly commands to move files, so it is easy to be rewritten into a bash script.
# However, I don't have a linux machine available now, hope someone can write one and open a PR.
cd builder
./build.bat
# install via pip
cd dist
pip install L0_Smoothing-*-py3-none-any.whlfrom L0_Smoothing import L0_Smoothing, L0_Smoothing_accel
# It is not recommended to use cv2.imread because it easily throws error
# just because you are missing some unimportant parameters and cannot
# read image from path with Chinese (and maybe other non-ascii) characters.
# Just read it with PIL and convert it to numpy array!
# Note that you need to convert image to BGR with cv2.cvtColor if you read with PIL.
import numpy as np
from PIL import Image
img = np.asarray(Image.open(r'/path/to/your/image'))
# Parameters:
# L0_Smoothing(img, asHSV=False, lambda_=2e-2, kappa=2.0, beta_max=1e5, mode='pyvkfft')
# L0_Smoothing_accel(img, asHSV=False, lambda_=2e-2, kappa=2.0, beta_max=1e5)
# img: numpy array of the image to be smoothed
# asHSV: This module does operation per channel, and you can choose to convert it to HSV
# while operating by giving parameter asHSV=True.
# lambda_, kappa, beta_max: read the paper
# mode: the OpenCL FFT backend to use
smoothed = L0_Smoothing_accel(img)If you are programming with pyopencl, you can use this module like this:
import pyopencl as cl
import numpy as np
import pyopencl.array as clArray
from L0_Smoothing import L0_Smoothing_CL
ctx = cl.create_some_context(interactive=False)
queue = cl.CommandQueue(ctx)
img = np.asarray(Image.open(r'/path/to/your/image'))
S = clArray.to_device(queue, img/255)
S_smoothed = L0_Smoothing_CL(S)
img = S.get()*255
img = np.clip(img, 0, 255).astype(np.uint8)Hint: You can try to apply blur before doing smoothing if the smoothing effect is not ideal.
Notes:
- It support only jpg and png images now.
- Input and output path should be both file or both folder.
- You can give lambda_, kappa, beta_max via
--params. - You can choose to use slower numpy version with switch
--noaccel. - If you give
showfor output path, processed image with not be saved but showed.
# get some help
L0-Smoothing --help
# process single image and save it somewhere
L0-Smoothing /path/to/your/image /path/you/want/to/save
# process all images in a folder and save it somewhere
L0-Smoothing /path/to/your/image/folder /path/you/want/to/save- When I use its command from terminal, it throws error!
Try adding switch--noaccelto use slower numpy version. - pip throws errors (on compiling) when I am installing pyvkfft!
Download pyvkfft package bundled with OpenCL SDK from release and install it with pip.
Or just download binary package from release and install it. - Binary packages have problem on my machine.
Usereiknaas fft backend, it is pure python.