Color Correction Meets Blind Validation for Image Capture: Are We Teaching to the Test?

It is common practice for digital image capture systems to use a small number of de-facto-standard test targets. Unfortunately, however, color (spectral-) characteristics of the colorants used may differ from those for the population of object/scenes to be captured. This can lead to poor color calibration of the system. A second limitation of current color-capture evaluation arises when the same set of color stimuli (color patches) are used to calibrate the color capture and to evaluate the residual color errors. When the same color-target is used, the reported color-encoding errors will usually be lower than those observed in normal image capture. This is because we are, in effect, ‘teaching to the test’, as when a student is prepared for test taking, rather than subject mastery. We can approach this under-reporting of color error as a measurement bias. We can treat color-correction (e.g. by a colorprofile) as being a statistical model relating the detected image values and their intended (‘correct’) pixel values. Using a statistical approach we adopt a validation method aimed at determining the extent to which this model relationship between variables (the regression model) provides an acceptable description of the data. For our color-imaging case, the equivalent step would be to test the computed color-correction (ICC profile) with color patches that are independent of those used to build the profile. We demonstrate a candidate strategy for selecting these test colors, and an example of a validation set of colors chosen to be distinct from the calibration set in the popular ColorChecker SG. Introduction Color image capture normally includes a color-correction step that transforms detector signals into corresponding pixel values. For digital cameras and scanners, we usually base the colorcorrection operation on captured images of reference color charts. From a colorimetric description of the input reference color patches (e.g. CIELAB coordinates, L*a*b*) and the corresponding (unprocessed) pixel values, we compute the color-correction parameters required for accurate color image encoding. This usually takes the form of either a custom or a popular output referred ICC profile. (e.g. sRGB, AdobeRGB, etc.) We can cast the building of an ICC color profile as a statistical modeling operation, where the model takes on the form specified by the profile elements, e.g., look-up tables, color matrix, etc. It is common practice for digital image capture systems to use a small number of de-facto-standard test target. Unfortunately, however, color (spectral-) characteristics of the colorants used may differ from those for the population of object/scenes to be captured. This can lead to poor color calibration of the system. The selection of collection-specific test targets for improved colorcapture has been addressed in the literature. In evaluating the goodness of any modeling there is normally a validation effort aimed at determining the extent to which the regression model provide an accurate description of the variables involved. A popular way to implement this regression is by way of an ICC color profile. In effect, this color profile acts as a color dictionary that translates triplets of captured RGB code values into equivalent colors as defined by the Profile Connection Space (PCS). The mathematical models to do so can be varied and sometimes complex. In the absence of a custom profile, often, standardized profiles are used. A popular way to evaluate the quality of this color calibration is to simply compare the translated color of each patch in PCS to the measured reference color of the actual target. While this is an instinctive approach, it yields, by definition, an optimal residual color error for that model since the regression model is designed to minimize such errors. One is effectively ‘teaching to the test’ when evaluating digital capture color performance using the same colors for which the color-correction was performed. We suggest using a validation approach where the color performance is tested with an independent and different set of color patches. Borrowing from medical clinical trials, these can be thought of as control (calibration) and treatment (validation) groups. While color calibrationor profiling validation is not often discussed in the literature, it can provide valuable information regarding the quality of image capture, and the likelihood of color artifacts during normal operation of the image capture system. Validation Color Patch Selection Set We recognize that the strategy for selecting a validation set of colors is open to infinite opinions. Being reasonable and without focusing on building a ‘killer’ validation target, we restrict our patch selection for validation using a set of criteria already included in the SG target. They are, • The same number of total patches • Identical set of gray patches (61) • Same number of chromatic patches (79) • Same number of patches within L*(10) slices • Semi-Gloss surface • Remained within the gamut of the existing CCSG The differences between the two sets are: • Different set of chromatic patches • Select patches from the Natural Color System (NCS) index We selected the chromatic patches by inspecting the CIELAB a*b* plots of each of eight L* slices (L*=10 increments). We identified gaps between the existing SG coordinates and selected an appropriate color from the NCS library of colors, as measured for this study. Figure 1 shows two example L* slices illustrating the SG calibration colors and the same number of NCS validation colors with that L* slice. Figure 2 illustrates a comparison image set between the SG calibration target and the validator target we will call SGX. The similarities and differences are consistent with the descriptions cited above. All of the patches, except for a few at L*>80, a*<20, and b* >90 (illuminant D50, 2 degrees) fit within the Adobe RGB color space. ©2016 Society for Imaging Science and Technology IS&T International Symposium on Electronic Imaging 2016 Image Quality and System Performance XIII IQSP-218.1 Figure 1: Comparison of calibration and validation CIELAB coordinates for two selected L* slices Experimental Custom colorimetric reference files were created for both targets. A raw digital image for both the Calibration and Validation targets was acquired (SG and SGX, respectively) from an Epson 10000XL scanner. Both were processed with a gamma 2.2 using all of the center 15 gray-patch values. Rough Profiler, built on the open-source Argyll colormanagement system, was used to create three different ICC profiles for each image using different methods (models). The models were labeled as, 1. Lab cLut, medium quality (Lcm) 2. Shaper + Matrix, medium quality ( SMm) 3. Lab cLut, high quality (Lch) At this point six different color profiles were now available. Each of the above three models for the two target images. The above profiles were then embedded into each of the two candidate image files and evaluated for color encoding accuracy. Figure 2: Comparison images of the SG (top) and SGX (bottom) targets Examples of the notation we will use to describe the image and profiling pairings are; • SGSG-LCM – SG image with a profile created using the SG target using a cLut medium quality profiling model. • SGXSGX-LCM SGX image with a profile created using the SGX target using a cLut medium quality profiling model. • SGXSG-LCM – SGX image with a profile created using the SG target using a cLut medium quality profiling model. The first two pairings above would be normal pairings of each target evaluated against a profile created by that target. It is the third pairing that is of interest where the validator target (SGX) is assessed for color encoding accuracy using a profile generated via the SG target. Results for several combinations of image target and ICC profile are presented in the Results section that follows. Results Table 1 lists the median and maximum ∆E2000 values for the important target-ICC profile combinations. For any particular statistical model, three sets of metrics are cited. Using the Lab cLut-medium quality model as an example, the SG-SG table numbers indicate lower median and maximum ∆E2000 values when the same IIC profile is used for the target from which it was derived. The same applies to the SGX-SGX combination. These two can act as baseline metrics for assessing the relative magnitude of the difference for the SGX-SG combination. This was the important validation combination for which this experiment was performed. The SGX-SG values were calculated using the SGX ©2016 Society for Imaging Science and Technology IS&T International Symposium on Electronic Imaging 2016 Image Quality and System Performance XIII IQSP-218.2 validation target but with an embedded ICC color profile calculated using the SG target. Using the median and maximum ∆E2000 values alone as indicators there does not appear to be a very large difference between the two. While there is a slight increase in encoding error by using the mismatched target-ICC profile validation combinations (i.e. SG-SGX combinations) it is not as large the authors expected. This applies to all three models. Indeed, the validation set for the shaper-matrix combination actually had a lower overall ∆E2000 compared to the target for which it was designed. Table 1 – Summary ∆E2000 for image-profile combinations ∆E2000 Target type Target profile source Model