Color Bregman TV

Michael Moeller, Eva-Maria Brinkmann, Martin Burger, Tamara Seybold

TL;DR

This paper introduces "Color Bregman TV," a novel iterative method for reconstructing multichannel data (like color images) by encouraging different channels to share a common edge structure. By minimizing the Bregman distance between channels across iterations, the algorithm effectively "borrows" edge information from one color channel to denoise another, leading to superior contrast and detail preservation compared to standard Total Variation (TV) methods.

What Problem Does This Paper Address?

In image processing, Total Variation (TV) regularization is a gold standard for removing noise while preserving edges. However, when dealing with color images, standard TV often treats the Red, Green, and Blue channels independently. This is wasteful because, in the real world, an edge (like the outline of an object) usually exists in all color channels simultaneously.

While "Vectorial TV" (VTV) exists to couple these channels, it often couples their intensities rather than their structural geometry. Furthermore, standard TV methods suffer from "contrast loss"—the reconstructed image often looks slightly washed out. The authors aim to solve two problems at once: enforcing structural consistency across color channels and restoring the contrast typically lost in variational reconstruction.

Methods

The authors generalize the Bregman iteration—originally a grayscale technique for restoring contrast—to the multichannel case.

The Core Algorithm

Instead of just regularizing a channel with its own TV functional, the authors propose an update rule where each channel is regularized by a weighted linear combination of Bregman distances to all other channels from the previous iteration.

Mathematically, the subgradient (which encodes edge location and direction) of one channel acts as a guide for the others. If the Red channel finds a strong edge, the Bregman distance term makes it "cheaper" for the Green and Blue channels to place an edge in that same spot.

Infimal Convolution

A clever refinement introduced is the use of infimal convolution of Bregman distances. Standard Bregman distances assume that if an edge exists in two channels, it must point in the same direction (e.g., both channels go from dark to light). However, in some images, one channel might go dark-to-light while another goes light-to-dark. The infimal convolution approach allows the model to enforce a common edge location while remaining independent of the edge direction (sign).

Implementation

The optimization problems are solved using the Split Bregman method (a variant of ADMM), which is computationally efficient and well-suited for L1-type problems like TV.

Datasets

The authors evaluated their method using:

  • Kodak Image Data Set: A standard benchmark of 24 high-quality color images.
  • McMaster (McM) Data Set: 18 images known for having high color saturation and complex structures.
  • Artificial Toy Examples: 1D signals and 2D "nested square" images used to specifically test edge alignment and sign independence.

Key Insights and Ideas

  • Subgradients as Edge Likelihoods: The "aha" moment is treating the subgradient of the TV functional as a "likelihood" map for edges. By sharing these subgradients across channels, the model performs a form of collaborative denoising.
  • Contrast Restoration: Because Bregman iterations effectively "add back the noise" (or rather, the residual) in a controlled way, they avoid the systematic bias toward flat, low-contrast regions seen in one-step TV minimization.
  • Structural vs. Intensity Coupling: Unlike VTV, which couples the values of the gradients, Color Bregman TV couples the structure (the normalized direction of the gradients). This makes it more robust for multimodal data where intensities might differ wildly but edges align.

Key Results

  • Denoising Performance: On the Kodak dataset, the Isotropic Color Bregman method achieved a mean PSNR of 32.95 dB, significantly outperforming standard Isotropic TV (31.59 dB) and Vectorial TV (32.16 dB).
  • Qualitative Superiority: Visual results (particularly on a bird's eye and wing) show that Color Bregman TV recovers fine details and dark spots that other methods completely smooth away.
  • Inpainting: In a "toy" inpainting task where different channels had different missing regions, the method successfully used information from intact channels to perfectly reconstruct the missing parts of others—a feat standard TV and VTV failed to achieve.

Limitations and Future Work

  • Theoretical Gaps: While the authors proved convergence for the basic Color Bregman iteration, a rigorous analysis of the "infimal convolution" version (including convergence speed and stationary points) is still lacking.
  • Computational Complexity: Iterative methods are inherently slower than one-step variational methods. While ADMM helps, the outer loop of Bregman iterations adds overhead.
  • Non-Local Extensions: The authors suggest that instead of just coupling color channels, one could couple similar patches within the same image (non-local coupling). This could allow a "clean" part of an image to help denoise a "noisy" but structurally similar part elsewhere.

Summary updated

Abstract

In this paper we present a novel iterative procedure for multichannel image and data reconstruction using Bregman distances. With the motivation that multiple channels sharing a common subgradient with respect to a suitable regularization implies desirable properties such as a common edge set (and a common direction of the normals to the level lines) in the case of the total variation (TV), we propose to determine each iterate by regularizing each channel with a weighted linear combination of Bregman distances to all other image channels from the previous iteration. In this sense we generalize the Bregman iteration proposed by Osher et al. to multichannel images. We prove the convergence of the proposed scheme, analyze stationary points and present numerical experiments on color image denoising, which show the superior behavior of our approach in comparison to TV, TV with Bregman iterations on each channel separately, and vectorial TV. Additionally, we propose to use the infimal convolution of Bregman distances to different channels from the previous iteration to obtain the independence of the sign and hence the independence of the direction of the edge. While this work focuses on TV regularization, the proposed scheme can potentially improve any variational multichannel reconstruction method with a 1-homogeneous regularization.
BibTeX
@article{moeller2014color,
  author = {Michael Moeller and Eva-Maria Brinkmann and Martin Burger and Tamara Seybold},
  title = {{Color Bregman TV}},
  journal = {SIAM J. Imaging Sci.},
  year = {2014},
  doi = {10.1137/130943388},
}
Figure