A new random color filter array with good spectral properties

Laurent Condat

TL;DR

This paper introduces a novel random color filter array (CFA) design that significantly reduces visually disturbing demosaicking artifacts compared to traditional periodic CFAs like the Bayer pattern. By arranging color filters in a non-periodic, blue-noise-like pattern, the authors achieve better spectral separation of luminance and chrominance information, leading to more pleasant visual results even with standard demosaicking algorithms.

What Problem Does This Paper Address?

Digital cameras capture color images using a sensor overlaid with a Color Filter Array (CFA). The most common is the Bayer CFA, which arranges red, green, and blue filters in a repeating pattern. To reconstruct a full-color image, a process called demosaicking is required to interpolate the missing color information for each pixel. A major challenge in demosaicking is aliasing, which arises from the undersampling inherent in the CFA process. In periodic CFAs like the Bayer pattern, aliasing often manifests as moiré patterns – structured, visually unpleasant artifacts. This paper addresses the problem of designing CFAs that minimize these moiré artifacts by distributing color information in a way that mimics the properties of human visual perception and reduces spectral overlap between luminance and chrominance signals.

Methods

The core contribution is a new method for generating random CFA patterns. The authors propose a "fast random tiling" algorithm that constructs the CFA in two steps:

  1. First Row and Column Generation: This step involves filling the first row and column of the mosaic using specific 3-pixel tiles (permutations of R, G, B). Tiles are chosen randomly from a set of admissible combinations based on the adjacent tile's color.
  2. Mosaic Filling: The rest of the mosaic is filled in scanline order. Each pixel is assigned a color different from its left and top neighbors. If there's a choice, the color that differs from the top-left diagonal neighbor is selected.

This method results in an aperiodic tiling where adjacent pixels always have different colors. The authors argue that this "blue noise" characteristic for chrominance channels concentrates their energy at higher frequencies, away from the baseband, thus reducing moiré artifacts.

Datasets

The paper evaluates its proposed CFA using a standard dataset of 20 color images, which were mosaicked and then demosaicked. No new datasets were introduced.

Key Insights and Ideas

  • Blue Noise for CFAs: The paper demonstrates that applying the principles of blue noise (low energy in the baseband, high energy at higher frequencies) to CFA design can effectively reduce demosaicking artifacts.
  • Randomness vs. Periodicity: Random, non-periodic patterns can offer advantages over periodic ones by avoiding the structured, moiré artifacts that arise from the periodicity of traditional CFAs.
  • Adjacent Pixel Color Constraint: The simple rule that adjacent pixels must have different colors is shown to be a powerful constraint for achieving desirable spectral properties in the CFA.
  • Qualitative Improvement: While quantitative improvements (MSE) are modest, the qualitative difference in artifact appearance (incoherent noise vs. moiré) is significant and visually more pleasing.

Key Results

  • The proposed random CFA, when used with standard demosaicking algorithms (bilinear, Lukac et al. [3], and a new linear method [15]), produces demosaicked images with artifacts that are "more visually pleasant" than those from the Bayer CFA.
  • In terms of Mean Square Error (MSE), the proposed CFA achieved competitive results, sometimes outperforming the Bayer CFA, especially with the new linear demosaicking method. For instance, with the proposed method [15], the MSE was 11.96, compared to 17.96 for the Bayer CFA.
  • Visual inspection of demosaicked images (e.g., a "lighthouse" image) clearly shows that the artifacts generated by the proposed CFA are less structured and more like random noise, whereas the Bayer CFA produces noticeable moiré fringes.

Limitations and Future Work

  • Intrinsic Balance: The authors acknowledge an intrinsic balance between concentrating energy around the Nyquist frequency (π) and avoiding energy around the zero frequency (0). Energy at π can correspond to local bi-colored checkerboards, which also have energy at 0.
  • Further Improvement: They believe there is still room for improvement in achieving better concentration of chrominance energy around π.
  • Beyond R,G,B: The paper notes that significant improvements have been shown with CFAs using colors other than R,G,B, but these often involve more complex physical realization. The current work focuses on R,G,B.
  • Practical Realization: While the proposed random pattern is appealing, the paper mentions that a periodically replicated small random pattern (e.g., 18x18) can be designed and used to tile the whole CFA, offering a practical approach that yields similar performance.

Summary updated

Abstract

We propose in this work a new random pattern with R,G,B colors, which can be used to design color filter arrays (CFA) with good spectral properties, since the chrominance is modulated at high frequencies. Hence, the demosaicking artifacts appear as incoherent noise, which is less visually disturbing than the moire¿ structures characteristic of periodic CFAs.
BibTeX
@inproceedings{condat2009new,
  author = {Laurent Condat},
  title = {{A new random color filter array with good spectral properties}},
  booktitle = {ICIP},
  year = {2009},
  doi = {10.1109/icip.2009.5413678},
}
Figure