How application-specific color calibration impacts multispectral imaging
Performance differences highlighted in tests of ink-jet samples
of paper and textile substrates
Timo Eckhard and Sebastian Georgi
Multispectral imaging cameras acquire various spectral image channels, typically more than the three channels acquired by conventional RGB cameras. Chosen for calibration testing was a Chromasens (Konstanz, Germany; www.chromasens.com) truePIXA multispectral camera, which acquires 12 distinct spectral image channels in the visible range of the light spectrum through a multi-channel CCD line sensor.
A color calibration algorithm then transforms the spectral image information to the desired output image format, such as a three-channel CIE-Lab image. This conversion is tailored to specific spectral properties of the material to be measured using the multispectral line-scan camera.
For certain groups of materials, color calibration can be generalized, meaning that using a color calibration on different types of material is possible. However, figuring out if this generalization is possible requires in-depth analysis. Instead, applying application-specific color calibration is recommended as this leads to improved multispectral imaging system performance, in terms of measurement capabilities.
Application-specific color calibration for multispectral color measurement in digital printing applications
In a test measurement analysis of ink-jet printed samples of paper and textile substrates a distinct, substrate-specific ink is used for each application. Due to the physical nature of matter and light interaction of ink and substrate, the spectral properties of a certain printed sample might be different for textile or paper, even though the color appears the same to the human observer. An example as such is given in Figure 1.
Color calibration of the test camera accounts for such spectral differences. To achieve this, a so-called “training chart” is used. For the ink-jet printed samples, a training chart may consist of a regular grid of printed color patches sampling the gamut of printable colors. (Figure 2). A Konica-Minolta (Tokyo, Japan; ) FD-7 spectroradiometer acquires spectral reference point measurements for each patch. Required for the color calibration apart from these spectral reference measurements is multispectral camera response data. This response data is extracted patch-wise from the regular grid by averaging image areas corresponding to the measurement spot of the spectroradiometric point-measurement device.
Using camera response and corresponding spectral reference measurements, a proprietary algorithm accessible in the truePIXA API/Chromantis graphical user interface (GUI) software from Chromasens performs color calibration.
In this test, color calibration is performed for three different substrates:
- Paper substrate
- Soft textile substrate with black backing
- Rough textile substrate with black backing
Figure 2 illustrates a small section of the training chart for each substrate.
As an example, the colorimetric and spectral properties of the patches included in the training chart for the case of paper substrate are shown in Figure 3 and Figure 4 respectively.
Denoted as “test charts,” a second set of color charts with differently-distributed color patches is produced on paper and textile substrates in the test, which enables color calibration and testing on distinct datasets.
The metrics considered for evaluation are spectral root mean square error (RMSE) and color difference in CIE-Lab color space (CIE-dE2000, further “dE00,” http://bit.ly/VSD-DE00), which is in accordance to the latest recommendation of the CIE. Using these metrics, differences between the measured patches and the reference measurement are computed.
In what follows, first order statistics of the metrics summarize the numerical results.
Findings are consistent for colorimetric and spectral metrics. However, for the sake of simplicity, interpretation is based mainly on color differences. Various conclusions can be drawn from the results.
In general measurement performance tests, high measurement performance is achieved when applying application-specific color calibration. In tests of measurement performance on paper vs. soft textile substrate, the results are comparable. Application-specific calibration is performed for each substrate, resulting in high-performance measurement results.
In tests of measurement performance on soft textile vs. rough textile substrate, the soft textile achieves higher performance as compared to rough textile. On average, dE00 is still smaller than one unit. The rough textile substrate is richer in texture. Judging visually, printed patches are spatially less homogeneous. It may potentially be that the reference measurement spot does not entirely match the measurement spot, which might explain the residual color differences.
For measurement performance testing on soft textile substrate with calibration for rough textile substrate, performance drops slightly compared to testing and calibration on rough textile substrate. On average, dE00 is still smaller than one unit and there is a considerable difference when interchanging the print substrate without applying a specific color calibration.
Lastly, in testing measurement performance on rough textile substrate with calibration for paper substrate, performance drops considerably compared to testing and calibration on paper substrate. The average color difference is so large that even a human observer could easily spot differences. Under this condition, multispectral color measurement does not result in adequate performance for industrial imaging applications.
Evaluations show measurement performance of multispectral image acquisition systems rely on color calibration being performed on an application-specific basis. For example, when considering textile substrates, color calibration must be performed with a textile substrate color calibration chart to achieve high measurement performance. Using paper substrate instead, color calibration has shown to considerably decrease performance. Therefore, performing application-specific color calibration when deploying multispectral imaging is recommended.
Performance drops off in these scenarios because of the intrinsic spectral properties of the different measurement substrates. Even though two objects from different material might appear similar in color—at least as humans perceive it—their spectral properties may be different.
Predicting a drop in measurement performance when using non-application specific color calibration is a safe bet. When application-specific color calibration isn’t feasible for technical reasons, handling these situations should be done on a case-by-case basis to evaluate the achievable measurement performance.