Sparse directional image representations using the discrete shearlet transform

Glenn R. Easley, Demetrio Labate, Wang‐Q LimApplied and Computational Harmonic Analysis 2007 · DOI

TL;DR

This paper introduces the Discrete Shearlet Transform (DST), a novel multiscale directional representation designed to overcome the limitations of traditional wavelets in capturing the geometric features of multidimensional signals like images, particularly edges. The DST offers optimal sparsity for representing images with edges and demonstrates competitive performance in denoising applications, both in terms of accuracy and computational efficiency.

What Problem Does This Paper Address?

Traditional wavelet transforms, while excellent for approximating signals with pointwise singularities, struggle with multidimensional data like images that often contain sharp discontinuities such as edges. Representing these edges accurately with wavelets requires a large number of coefficients, leading to inefficient representations. This paper addresses the need for a representation system that is not only multiscale but also possesses high directional sensitivity to effectively capture the geometric structures in images, thereby achieving optimal sparsity and improved performance in applications like denoising.

Methods

The core contribution is the Discrete Shearlet Transform (DST). This transform is based on the shearlet transform, which combines multiscale analysis with a unique ability to capture geometric features. The paper presents two implementation methods:

  • Frequency-Domain Implementation: This approach leverages the Laplacian pyramid algorithm and a pseudo-polar grid for directional filtering. It offers flexibility in choosing window functions and allows for subsampling, making it suitable for compression.
  • Time-Domain Implementation: This method uses a direct convolution approach with shearing filters. It is designed for applications requiring translation invariance, such as denoising, and aims to reduce Gibbs-type ringing by using filters with smaller support sizes.

The mathematical framework of shearlets is built upon affine systems with composite dilations, allowing for representations that capture scale, location, and orientation.

Datasets

The paper uses standard test images for evaluation, including:

  • Lena
  • Peppers
  • Goldhill
  • Elaine
  • Cameraman

These images are used to demonstrate the performance of the shearlet transform in denoising and approximation. No new datasets are introduced.

Key Insights and Ideas

  • Superior Edge Representation: Shearlets are inherently better at capturing directional information, especially edges, compared to wavelets. This leads to sparser representations for images with these features.
  • Optimal Sparsity: The shearlet transform achieves an optimal sparsity rate for representing images with edges, theoretically approaching $O(N^{-2} (\log N)^3)$ for the approximation error with $N$ coefficients.
  • Flexibility in Implementation: The paper presents both frequency-domain and time-domain implementations, offering different advantages for various applications (e.g., compression vs. denoising).
  • Computational Efficiency: Both implementations are shown to be computationally efficient, with the time-domain approach offering faster inversion and reduced artifacts due to smaller filter support.
  • Competitive Denoising Performance: The DST, particularly the time-domain version, demonstrates superior denoising performance compared to wavelets, contourlets, and other advanced methods, achieving higher PSNR values.

Key Results

  • Denoising Performance: In denoising experiments, the time-domain shearlet transform (NSST1(16,32)) consistently outperformed other methods, including wavelets, contourlets, and BivShrink, by a significant margin (often 1 dB or more in PSNR). For example, on the "Lena" image with $\sigma=10$, NSST1(16,32) achieved 35.38 dB, compared to 35.29 dB for NSCT and 35.36 dB for DDWT.
  • Approximation Error: Numerical experiments confirmed the theoretical predictions of optimal approximation rates, showing decay rates close to $M^{-1}$ for images with smooth regions and smooth edges.
  • Computational Efficiency: The time-domain shearlet transform demonstrated competitive CPU times, with inversion being particularly fast. For a 512x512 image, SFC with 16 directions took 2.9266 x 10⁻² sec, and iSFC took 1.0938 x 10⁻¹ sec.

Limitations and Future Work

  • Anisotropic Subsampling: The frequency-domain implementation uses downsampling in vertical and horizontal directions, leading to redundancy. The paper notes that incorporating anisotropic subsampling is a non-trivial matter for future investigation.
  • Filter Support: While the time-domain filters are designed to have smaller support to reduce artifacts, they are not compactly supported in the traditional sense.
  • Future Research: The authors intend to explore other applications of the discrete shearlet transform, such as image deconvolution, and study its uses in other multidimensional data applications. They also suggest investigating alternative discretizations of the Meyer wavelet to mitigate performance degradation in frequency-domain implementations.

Summary updated

BibTeX
@inproceedings{easley2007sparse,
  author = {Glenn R. Easley and Demetrio Labate and Wang‐Q Lim},
  title = {{Sparse directional image representations using the discrete shearlet transform}},
  booktitle = {Applied and Computational Harmonic Analysis},
  year = {2007},
  doi = {10.1016/j.acha.2007.09.003},
}
Figure