Spectral characterization of a color scanner by adaptive estimation

Hui‐Liang Shen, John H. XinJournal of the Optical Society of America A 2004 · DOI

TL;DR

This paper introduces a novel "adaptive estimation" (AE) method for accurately characterizing color scanners. It tackles the challenge of scanners deviating from linear behavior by adaptively learning local color statistics, leading to significantly improved spectral reflectance estimation compared to existing linear models.

What Problem Does This Paper Address?

Accurately capturing color information is crucial for digital imaging. Color scanners, the devices that digitize physical objects, have inherent limitations. To faithfully reproduce colors across different devices, we need to "characterize" them – essentially, understand how they translate physical colors into digital signals. This can be done colorimetrically (mapping to device-independent color values) or spectrally (recovering the actual spectral reflectance of the object).

Spectral characterization is more robust as it's independent of lighting conditions. However, a key challenge is estimating the scanner's spectral responsivity (how its sensors respond to different wavelengths). This is difficult to measure directly and often relies on mathematical recovery. Existing methods often assume the scanner behaves linearly, but real scanners can deviate from this. When the recovered spectral responsivity is inaccurate due to non-linearity, these methods struggle. This paper addresses this by proposing a method that doesn't strictly rely on an accurate, pre-recovered spectral responsivity, making it more robust to scanner non-linearities.

Methods

The paper proposes an Adaptive Estimation (AE) method. The core idea is to move beyond a single, global model of the scanner's behavior and instead learn local models based on the specific color being scanned.

  1. Scanner Linearity Investigation: First, the authors investigate the scanner's linearity. They use a constrained, linear, least-squares method to recover the scanner's spectral responsivity. They also model the non-linear input-output behavior of the scanner using an optoelectronic conversion function.
  2. Global Transform Estimation: A global transform matrix is calculated based on a minimum-mean-square-error criterion. This provides a baseline, non-adaptive model.
  3. Adaptive Estimation (AE): This is the novel part. For a candidate color sample:
  4. Its estimated reflectance is calculated using the global model.
  5. The algorithm then searches for L neighboring training samples in the reflectance space that are closest to this estimated reflectance.
  6. An adaptive transform matrix is then calculated using only these L neighboring samples. This local model is expected to better capture the scanner's behavior for that specific color.
  7. The final reflectance is estimated using this adaptive matrix.

The key advantage is that the AE method's performance is not critically dependent on the accuracy of the initially recovered spectral responsivity, as it focuses on local statistics.

Datasets

The paper uses the following color targets for training and evaluation:

  • GretagMacBeth ColorChecker Chart (MCC): Used for investigating scanner linearity and recovering spectral responsivity.
  • GretagMacBeth ColorChecker DC (CDC): Used for evaluating the proposed AE method.
  • Kodak Gray Scale Q-14 (Q14): Used for obtaining the nonlinear function.
  • Kodak Q60 photographic standard (IT8): Used for evaluating the proposed AE method. Reflectance data for IT8 was obtained from an online repository.

The authors note that they clipped the IT8 reflectance data to match the spectral range of the other targets (400-700 nm).

Key Insights and Ideas

  • Scanner non-linearity is a significant challenge: The paper highlights that real scanners often deviate from linear behavior, which can severely impact the accuracy of spectral characterization methods that rely on accurate spectral responsivity.
  • Local statistics are key for robustness: By adaptively estimating the scanner's behavior based on neighboring color samples, the AE method becomes more resilient to the global non-linearities and inaccuracies in the recovered spectral responsivity.
  • The "candidate sample itself is not included" in AE training: This is an important detail for realistic characterization, as usually, the exact sample being characterized isn't part of the training set.
  • The optimal number of neighbors (L) is color-target related: The paper demonstrates that the ideal number of neighboring samples for adaptive estimation varies depending on the specific color target used.

Key Results

  • Significant advantage of AE: Experimental results show that the proposed AE method significantly outperforms existing methods like the Linear Reflectance Model (LRM) and Non-Adaptive Estimation (NAE) in terms of both color difference (ΔE⁹⁴) and root-mean-square (rms) error in reflectance estimation.
  • Improved accuracy: For the CDC target, the AE method achieved a mean color difference of 1.63 ΔE⁹⁴ units and an rms error of 0.0179. For the IT8 target, these values were 1.44 ΔE⁹⁴ units and 0.0066, respectively. These are considered sufficient for many applications.
  • LRM limitations: The LRM method, especially when the candidate sample is excluded from the training set (which is more realistic), showed considerably higher errors compared to AE.
  • Determining optimal L: For the CDC target, L=20 was found to be appropriate, while for IT8, L=10 was suitable.

Limitations and Future Work

  • Data dependency of L: The optimal number of neighboring samples (L) for adaptive estimation is found to be color-target related. This suggests that determining L might require some prior knowledge or an additional optimization step for new targets.
  • Computational cost: While not explicitly detailed as a limitation, adaptive estimation inherently involves searching for neighbors, which can be computationally more intensive than a single global model.
  • Future work: The paper suggests that the AE method has potential applications in areas requiring high color accuracy, such as textile quality control and art painting recording. Further exploration of optimal L selection and computational efficiency could be future directions.

Summary updated

Abstract

A new method of spectral characterization for color scanners by the use of adaptive estimation is proposed. It deals with estimation of high-dimensional reflectance vectors from low-dimensional scanner response vectors when the scanner departs from linearity. We first investigate the spectral linearity of the scanner, and then estimate the spectral reflectance adaptively based on the local statistics of a set of neighboring training samples. As the proposed characterization method does not utilize the mathematically recovered spectral responsivity, its inherent inaccuracy is not critical to the spectral characterization. Experimental results showed significant advantage of adaptive estimation when compared with other methods.
BibTeX
@article{shen2004spectral,
  author = {Hui‐Liang Shen and John H. Xin},
  title = {{Spectral characterization of a color scanner by adaptive estimation}},
  journal = {Journal of the Optical Society of America A},
  year = {2004},
  doi = {10.1364/josaa.21.001125},
}
Figure