2D Dual-Tree Complex Biorthogonal M-Band Wavelet Transform

Caroline Chaux, Jean-Christophe Pesquet, Laurent Duval

TL;DR

This paper introduces a novel extension of the dual-tree complex wavelet transform to M-band biorthogonal filter banks. This new framework offers greater flexibility in directional image analysis by allowing for more choices in wavelet design and increasing the number of selectable directions, leading to improved denoising performance.

What Problem Does This Paper Address?

Analyzing natural images often requires sophisticated tools that can capture directional features and provide efficient representations. While traditional wavelet transforms are powerful, they can suffer from limitations like shift-variance and restricted directional selectivity. The dual-tree complex wavelet transform (DT-CWT) was developed to address these issues, offering improved shift-invariance and directional analysis. Previous work extended the DT-CWT to M-band (using M filters instead of 2) for real wavelets, increasing flexibility. However, this paper tackles the problem of further enhancing this framework by incorporating two key aspects: biorthogonality and complex M-band filter banks. Biorthogonality allows for more freedom in designing wavelets with desirable properties (like symmetry and compact support), while the complex M-band extension promises even richer directional analysis capabilities.

Methods

The authors extend the M-band dual-tree wavelet transform by focusing on two main fronts: biorthogonal wavelets and complex M-band filter banks.

  1. Biorthogonal Hilbert Pairs of Wavelets: They first address the construction of biorthogonal wavelet pairs in the M-band case. This involves designing analysis and synthesis filter banks that satisfy specific reconstruction properties. A key technique is defining Hilbert pairs of wavelets using relationships involving the signum function and phase functions, which are then used to construct the biorthogonal filters.
  2. Complex 2D Dual-Tree M-Band Wavelet Transform: The core contribution is the extension of the complex DT-CWT to the M-band scenario. This is achieved by leveraging complex-valued wavelets that can be either analytic or anti-analytic. By combining these in tensor products, the authors demonstrate how to select specific quadrants of the frequency plane for directional analysis. This leads to the need for four separable 2D wavelet analyses.
  3. Direction Selection: The paper details how different combinations of analytic/anti-analytic wavelets (controlled by parameters $\epsilon_1, \epsilon_2$) allow for the selection of specific directional subbands in the 2D frequency plane. Figure 1 visually illustrates this concept for an M=4 case.
  4. Prefiltering and Reconstruction: For digital implementation, prefiltering stages are introduced to transition from the analog formalism to discrete processing. The reconstruction process is also discussed, acknowledging the redundancy introduced by the complex dual-tree transform and proposing a minimization problem to find a robust reconstruction.

Datasets

The paper evaluates its proposed method using the "Barbara" image, a standard 512x512 image commonly used for image processing research, corrupted with additive white Gaussian noise. No new datasets were introduced.

Key Insights and Ideas

  • Enhanced Directional Selectivity: The complex M-band extension significantly increases the number of selectable directions compared to previous real M-band or dyadic complex DT-CWTs. This is crucial for analyzing complex image structures.
  • Flexibility through Biorthogonality: Incorporating biorthogonal wavelets provides greater freedom in designing wavelets with desirable properties (e.g., symmetry, compact support) that are often difficult to achieve with orthogonal wavelets, especially in M-band settings.
  • Trade-off between Orthogonality and Biorthogonality: The results suggest that while orthogonal decompositions offer strong performance, biorthogonal decompositions (specifically the 9/7 wavelets) yield comparable results and might be more advantageous for image coding applications.
  • M-band Advantage: The study reinforces the benefit of M-band filter banks over dyadic ones, consistently showing improvements in denoising performance across various estimators.

Key Results

The paper demonstrates the effectiveness of the proposed complex M-band biorthogonal dual-tree wavelet transform through image denoising experiments on the "Barbara" image.

  • SNR Improvements: Table 1 shows denoising results in terms of Signal-to-Noise Ratio (SNR) in dB. For instance, the complex 4-band biorthogonal transform (DTTbi4cx) achieves an SNR of 10.90 dB using the bivariate shrinkage estimator, which is a significant improvement over the initial SNR of 5.67 dB.
  • M-band vs. Dyadic: All 4-band structures consistently outperform their dyadic counterparts, offering improvements of at least 0.2 dB.
  • Complex vs. Real: Complex decompositions generally outperform real ones, though the difference is smaller for bivariate shrinkage.
  • Biorthogonal Performance: The 9/7 biorthogonal decompositions provide results close to orthogonal ones, highlighting their competitive performance.

Limitations and Future Work

  • Redundancy: The complex dual-tree transform introduces a redundancy factor of 4, which, while enabling better analysis, requires careful consideration during reconstruction. The paper addresses this by proposing a minimization problem for reconstruction.
  • Computational Cost: While the paper mentions reduced computational cost for the standard DT-CWT, the M-band complex extension likely increases this cost due to the increased number of filters and subbands. This is not explicitly quantified.
  • Future Work: The authors suggest that biorthogonal decompositions could be more useful in image coding applications, implying that further research in this area is warranted. They also mention that the proposed framework could be applied to other image processing tasks beyond denoising.

Summary updated

Abstract

Dual-tree wavelet transforms have recently gained popularity since they provide low-redundancy directional analyses of images. In our recent work, dyadic real dual-tree decompositions have been extended to the M-band case, so adding much flexibility to this analysis tool. In this work, we propose to further extend this framework on two fronts by considering (i) biorthogonal and (ii) complex M-band dual-tree decompositions. Denoising results are finally provided to demonstrate the validity of the proposed design rules.
BibTeX
@inproceedings{chaux2007d,
  author = {Caroline Chaux and Jean-Christophe Pesquet and Laurent Duval},
  title = {{2D Dual-Tree Complex Biorthogonal M-Band Wavelet Transform}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2007},
  url = {https://openreview.net/forum?id=MYJCODHzV8},
}
Figure