Analysis of Singularities and Edge Detection using the Shearlet Transform

Glenn R. Easley, Kanghui Guo, Demetrio LabateHAL 2009

TL;DR

This paper introduces an improved edge detection algorithm for images by leveraging the shearlet transform. Unlike traditional wavelets, the shearlet transform excels at capturing both the location and orientation of discontinuities, leading to more robust edge detection, especially in noisy images.

What Problem Does This Paper Address?

Detecting edges in images is a fundamental task in computer vision and image analysis. Traditional methods often struggle with noise and accurately characterizing the geometric properties of edges, such as their orientation and curvature. While wavelets offer multiscale analysis and can signal the presence of discontinuities, they lack the directional sensitivity needed for precise edge localization and characterization. This paper addresses this gap by proposing a shearlet-based approach that can better capture the geometric information of edges, leading to more accurate and robust edge detection.

Methods

The core of this paper is the application of the continuous shearlet transform for edge detection. The shearlet transform is a multiscale and directional representation that extends the concepts of wavelets and curvelets. It decomposes a signal into components localized in scale, location, and orientation.

Key aspects of the method include:

  • Shearlet Transform Definition: The paper defines the shearlet transform using anisotropic dilations and shears, allowing it to adapt to the directional nature of image features.
  • Asymptotic Decay Property: A crucial theoretical result (Theorem 2.1) highlights that the shearlet transform exhibits rapid asymptotic decay everywhere except at edge locations and for orientations normal to those edges. This property is key to identifying edges.
  • Lipschitz Regularity: The paper connects edge detection to the concept of Lipschitz regularity, where edges are characterized by a Lipschitz exponent of 0. The shearlet transform's ability to precisely characterize singularities is linked to this.
  • Discrete Shearlet Transform: While the theoretical foundation is based on the continuous transform, the paper mentions that a discrete version was developed and applied for practical edge detection.
  • Edge Detection Algorithm: The shearlet transform is used to analyze the image at multiple scales and orientations. The regions with significant shearlet coefficients, particularly those aligned with edge orientations, are identified as edges.

Datasets

The paper uses standard image datasets for evaluation. Specifically, it presents results on:

  • A brain MRI image.
  • A retinal blood vessel image.

These images are used to compare the performance of the proposed shearlet-based edge detector against traditional methods.

Key Insights and Ideas

  • Directional Sensitivity is Crucial: The paper strongly emphasizes that for accurate edge detection, especially in the presence of noise, capturing the orientation of discontinuities is as important as their location and scale.
  • Shearlets Excel at Geometric Characterization: The shearlet transform's ability to provide a precise geometrical characterization of discontinuities, including irregularities like corners and junctions, makes it superior to wavelets for this task.
  • Asymptotic Decay as an Edge Indicator: The theoretical finding that the shearlet transform decays rapidly everywhere except at edges provides a powerful mathematical basis for edge detection.
  • Robustness to Noise: The shearlet-based approach demonstrates improved robustness to noise compared to wavelet-based and traditional methods, as evidenced by the experimental results.

Key Results

The paper presents qualitative and quantitative results comparing the shearlet-based edge detection with wavelet-based, Sobel, and Prewitt methods.

  • Qualitative Improvement: Figures 2 and 3 show that the shearlet-based method produces cleaner and more continuous edge maps, with fewer spurious edges and better preservation of edge structure, especially in noisy images.
  • Quantitative Improvement (Pratt's Figure of Merit - FOM):
    • For the MRI image (PSNR=28.10 dB): Shearlet (FOM=0.44) significantly outperforms Sobel (FOM=0.24), Wavelet (FOM=0.29), and Prewitt (FOM=0.23).
  • For the retinal image (PSNR=24.58 dB): Shearlet (FOM=0.45) again outperforms Sobel (FOM=0.15), Wavelet (FOM=0.27), and Prewitt (FOM=0.15).

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, the paper mentions that a discrete version of the shearlet transform was developed, implying that the continuous transform might be computationally intensive. Future work likely involves optimizing the discrete implementation.
  • Parameter Tuning: The paper states that "consistent set of predetermined default parameters" were used for comparison, suggesting that parameter tuning could further optimize performance.
  • Further Applications: The authors hint at the broader applicability of the shearlet transform for analyzing singularities and suggest that the presented edge detection algorithms are an example of its utility. Future work could explore other applications in image analysis and signal processing where directional multiscale analysis is beneficial.

Summary updated

Abstract

The continuous curvelet and shearlet transforms have recently been shown to be much more effective than the traditional wavelet transform in dealing with the set of discontinuities of functions and distributions. In particular, the continuous shearlet transform has the ability to provide a very precise geometrical characterization of general discontinuity curves occurring in images. In this paper, we show that these properties are useful to design improved algorithms for the analysis and detection of edges. 1.
BibTeX
@inproceedings{easley2009analysis,
  author = {Glenn R. Easley and Kanghui Guo and Demetrio Labate},
  title = {{Analysis of Singularities and Edge Detection using the Shearlet Transform}},
  booktitle = {HAL},
  year = {2009},
}
Figure