Shearlets

Gitta Kutyniok, Demetrio Labate, John Benedetto, Stephan Dahlke, Glenn Easley, Philipp Grohs, Kanghui Guo, Jakob Lemvig, Tomas Sauer, Gabriele Steidl, Gerd TeschkeApplied and Numerical Harmonic Analysis 2012 · DOI · OpenAlex

Abstract

Fourier transforms are used, in conjunction with the uncertainty principle, for dealing with signal reconstruction in communications theory. This work returns to the raison d'être of the Applied and Numerical Harmonic Analysis (ANHA) series, exploring advanced multiscale frameworks for multivariate data.

Introduction: Beyond Isotropic Wavelets

While wavelets revolutionized signal processing for 1D data in the 1990s, they struggle to efficiently represent "edges" and "curves" in higher dimensions. The primary limitation is that wavelets are isotropic—they scale equally in all directions. This makes them excellent at catching point-like singularities but inefficient for lines or surfaces.

In a 2D image, an edge is an anisotropic feature with a specific position and orientation. Representing a smooth curve with wavelets requires a massive number of coefficients because the wavelets do not "align" with the geometry. Shearlets address this by providing a mathematical framework capable of handling directional features with optimal sparsity, using the fewest possible pieces of information to represent complex geometric structures.

The ANHA series encompasses a wide range of signal processing and mathematical topics, providing the context for the development of Shearlets.

The Shearlet Transform

The core innovation is the Shearlet Transform. Unlike previous attempts to fix wavelets, such as Curvelets (which rely on rotation), Shearlets use a combination of three operations:

  1. Scaling: Adjusting the size to analyze different frequencies (multiscale).
  2. Translation: Moving the window to different spatial locations (localization).
  3. Shearing: Using a shearing matrix to change the orientation.

From a computational standpoint, rotation is difficult to implement on a square pixel grid without losing data or introducing artifacts. Shearing, however, preserves the integer grid, facilitating a "faithful" transition from continuous mathematical theory to discrete digital algorithms. The framework supports both band-limited shearlets for theoretical precision and compactly supported shearlets for high-speed local processing.

Key Insights and Applications

Shearlets are proven to achieve the theoretical limit for approximating "cartoon-like" images—piecewise smooth functions separated by smooth curves. They represent edges far more efficiently than wavelets or Fourier transforms.

Microlocal Analysis and 3D Data

One significant advantage is that shearlets can precisely characterize the orientation and "wavefront set" of a singularity. Furthermore, because the shearing operator maps the digital grid to itself, the algorithms avoid the "leakage" common in other directional transforms. Shearlets are also uniquely suited for 3D data; while 2D edges are curves, 3D edges are surfaces, and the shearing operation scales naturally to these higher dimensions.

Practical Performance

In applications like image denoising, deconvolution, and inpainting, shearlet-based algorithms (specifically those in the ShearLab library) consistently outperform traditional wavelet methods in both Peak Signal-to-Noise Ratio (PSNR) and visual edge preservation. In medical imaging, the framework shows promise in inverting the Radon transform, potentially leading to clearer CT scans with lower radiation doses.

Limitations and Future Work

Despite their advantages, the directional sensitivity of shearlets requires more memory and processing power than simple 1D wavelets. Many current shearlet transforms are also "redundant," producing more data than the original signal. Future research focuses on developing "critically sampled" versions that maintain directional sensitivity without excess data. Additionally, moving these complex multiscale analyses into real-time video processing or mobile hardware remains a significant engineering challenge.

Summary updated

Abstract

Fourier transforms, and these are used, in conjunctionwith the uncertainty principle, for dealing with signal reconstruction in communications theory.We are back to the raison d'tre of the ANHA series!
BibTeX
@inproceedings{kutyniok2012shearlets,
  author = {Gitta Kutyniok and Demetrio Labate and John Benedetto and Stephan Dahlke and Glenn Easley and Philipp Grohs and Kanghui Guo and Jakob Lemvig and Tomas Sauer and Gabriele Steidl and Gerd Teschke},
  title = {{Shearlets}},
  booktitle = {Applied and Numerical Harmonic Analysis},
  year = {2012},
  doi = {10.1007/978-0-8176-8316-0},
}
Figure