The weighting function for lightness in the CIE94 color-difference model

Abstract

We have tested the potential improvement achieved by the CIE94 color‐difference model [CIE Publication 116–1995] when a linear function of lightness (SL = K1 + K2 L*) is introduced as a weighting function for the lightness difference ΔL*. Our analyses are based on experimental results previously obtained using object colors and visual colorimeters; new experimental results are not reported here. For the RIT‐Dupont dataset [Color Res. Appl. 16, 297–316, 1991], the optimal values of the linear function tested are K1 = 1.02; K2 = 0.001, in excellent agreement with the proposal of the CIE94 model: K1 = 1.0; K2 = 0. However, for the dataset of Coates et al. [J. Soc. Dyers Colour. 97, 179–183, 1981], the best predictions were obtained using K1 = 0.5; K2 = 0.008, improving the results obtained by CMC(l:c) or BFD(l:c). These last values differ by a factor two from the proposal K1 = 1.0; K2 = 0.017 made in another recent study [Color Res. Appl. 19, 273–276, 1994]. Parametric factors, such as the texture of the samples and the lightness of the background, appear to be the main causes of the discrepancies found between the best weighting function for lightness, using different experimental datasets. While new experimental results are reported to clarify this problem, the present results support the proposal made by the CIE94 model, suggesting that the lightness‐dependence correction can be considered as nonrobust and related to local parametric effects. © 1996 John Wiley & Sons, Inc.

Revisiting the weighting function for lightness in the CIEDE2000 colour‐difference formula

“…[28], BIGC-T2-M [29], Wang [31], and BIGC-S-SG [32]), and V-inverse-shaped functions for two datasets (BIGC-T2-SG [29] and BIGC-T2-G [29]). Results for each one of the 13 visual datasets (see the key) can be distinguished trends in these three datasets are unknown, the possibility can be suggested that different materials or textures of samples (see Table 2) are a potential cause of the discrepancies leading to the proposal of very different S L functions in the literature [5,6,13,18,22,[23][24][25]. Displacements (in different senses) in the minima of the fitted V-shaped functions with respect to the value 50 are also appreciable for the BFD-P and BIGC-T2-M datasets.…”

“…While the reasons for the different (2) is modified at a time by a percentage of AE100%. Results for each one of the 13 visual datasets (see the key) can be distinguished trends in these three datasets are unknown, the possibility can be suggested that different materials or textures of samples (see Table 2) are a potential cause of the discrepancies leading to the proposal of very different S L functions in the literature [5,6,13,18,22,[23][24][25]. It must be recognised that very few curves in Figure 3 closely resemble the prediction made by CIEDE2000, and we must admit that a clear lightness crispening may not be occurring in all datasets considered in Figure 3.…”

“…This S L function is a symmetrical curve with a minimum at L* = 50, corresponding to the assumed lightness of the background behind the samples in the colour pairs. These proposals are related to divergent results found in different experiments [23][24][25]. Thus, the CIEDE2000 colour-difference formula predicts a minimum tolerance when the colour pairs have the same lightness as the background (which is called the 'lightness crispening effect'), and higher tolerances for samples darker or lighter than the background.…”

“…For example, the CMC and BFD colour‐difference formulas proposed small/high tolerances for dark/light samples, the LCD formula proposed higher tolerances only for samples lighter than the background, and the CIE94 formula proposed independence of tolerance with the lightness of the samples. These proposals are related to divergent results found in different experiments . Currently, the controversy concerning the optimal S L function remains open.…”

<title>Issues of measurement and assessment in hard copy color reproduction</title>

“…C1E94 S non-Euclidean, since the way in which the formula is defined leads to a difference in \E according to which of the two samples is taken as the reference. For some high chroma sample pairs (which may have an acceptable difference with C*ab chroma differences as high as 12), this can lead to a difference of about O.5AE depending upon which sample is taken as the reference. We assume that this is why the CIE recommendation is that C1E94 should only be used if the CIELAB LE is less than 5.…”