A New Contourlet Transform with Sharp Frequency Localization

[Paper] This is the PyTorch implementation of ContourletSD: A New Contourlet Transform with Sharp Frequency Localization

Yue M. Lu¹, Minh N. Do^2
1Harvard University, ²University of Illinois at Urbana-Champaign

Dependencies.

For a full list of requirements, please refer to requirements.yml. To install the dependencies, install mini-conda and execute:

conda env create -f requirements.yml

Getting Started.

Our contourlets implementation is based on native PyTorch and allows fast access to directional multiresolution image representations.

The ContourletSD library and helper functions can be imported as:

import torch
from torchvision.io import read_image

from contourletSD_pytorch.contourlet_ops import (get_norm_scaling, hard_thresholding)
from contourletSD_pytorch.contourlet_sd import ContourletSD
from utils.smooth_functions import VALID_SMOOTH_FUNC

The contourlet analysis and synthesis operators can be used via the ContourletSD class:

contourlet_sd = ContourletSD(
    nlevs=[2, 2, 3, 4, 5],
    Pyr_mode=1,
    smooth_func=VALID_SMOOTH_FUNC['rcos'],
    dfilt='pkva',
    color_mode='rgb',
)

After reading an image and generating its noisy version:

X = read_image('datasets/test_images/peppers_color.png').unsqueeze(0).to(torch.float)
Xn = X + 30.0 * torch.randn_like(X)

Denoising via contourlet hard thresholding can be performed as:

# Get contourlet coefficients.
Y = contourlet_sd(x=Xn)

# Load pre-computed norm scaling factors for each subband.
E = get_norm_scaling(
    image_size=Xn.shape[-1],
    SDmm_dir='extras/sdmm_matlab',
    Pyr_mode=1,
    device=Xn.device,
)

# Apply hard thresholding on coefficients.
Yth = hard_thresholding(
    y=Y,
    sigma=30.0,
    E=E,
)

# Reconstruct image.
Xd = contourlet_sd(
    x=Yth,
    reconstruct=True,
)

Results:

Demos.

The scripts subfolder includes bash scripts to run denoising demos on CPU and CUDA if available. To run the CUDA demos, the Set12 and Color Berkeley Segmentation (BSDS300) datasets must be included in the datasets subfolder.

References.

If you use this code or find it helpful, please consider citing:

@inproceedings{lu_2006_new,
  title={A new contourlet transform with sharp frequency localization},
  author={Lu, Yue and Do, Minh N},
  booktitle={2006 International Conference on Image Processing},
  pages={1629--1632},
  year={2006},
  organization={IEEE}
}