L0-Smoothing 0.1.3

Implementation of 《Image Smoothing via L0 Gradient Minimization》

Image Smoothing via L0 Gradient Minimization

Blog Post | Report Bug on Numpy Version | Report Bug on PyOpenCL Version

Table of Contents

• About the Project
• Getting Started
  • Installation
    • Install from pypi
    • Setup with source code
  • Usage
    • Import in your script
    • Execute from terminal
• Maybe FAQ

About the Project

This repository is the Python implementation of the paper: Image Smoothing via L0 Gradient Minimization

Flower

Rock

Getting Started

Installation

Install from pypi

pip install L0-Smoothing

Setup with source code

# 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.whl

Usage

Import in your script

from 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.

Execute from terminal

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 show for 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

Maybe FAQ

• When I use its command from terminal, it throws error!
  Try adding switch --noaccel to 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.
  Use reikna as fft backend, it is pure python.