Comparison of Colour Difference Methods for Natural Images

Perceptual colour difference in simple colour patches has been extensively studied in the history of colour science. However, these methods are not assumed to be applicable for predicting the perceived colour difference in complex colour patches such as digital images of complex scene. In this work existing metrics that predict the perceived colour difference in digital images of complex scene are studied and compared. Performance evaluation was based on the correlations between values of the metrics and results of subjective tests that were done as a pair comparison, in which fifteen test participants evaluated the subjective colour differences in digital images. The test image set consisted of eight images each having four versions of distortion generated by applying different ICC profiles. According to results, none of the metrics were able to predict the perceived colour difference in every test image. The results of iCAM metric had the highest average correlation for all images. However, the scatter of the judgements was very high for two of the images, and if these were excluded from the comparison the Hue-angle was the best performing metric. It was also noteworthy that the performance of the CIELAB colour difference metric was relatively high. Introduction The conventional CIE metrics (e.g. CIEDE2000 /1/)) developed to estimate the colour differences of colour fields are capable to achieve a degree of prediction that is commonly acknowledged to be sufficient. These metrics require that the two stimuli being matched are presented using identical backgrounds and surroundings, and also that the two stimuli are viewed using identical illuminants and observers defined by the CIE. The results of the metrics are unreliable when these requirements are not met. Furthermore, these metrics are being used in quality control of colour reproduction, in which the recent cross media demands have made this conventional colorimetry insufficient. The CIECAM97 model and the updated CIECAM02 model were developed to provide a viewing condition specific method for transforming tristimulus values into perceptual attribute correlates /2/. However, these models can only interpret simple colour patches due to their nonexistent capabilities to model the properties of spatial structure in complex images. These properties have received considerable attention in different fields of colour science, such as study of image similarity and retrieval, image segmentation, image quality and human colour vision. The definition of complex images rises from their structure, which consists of different spatial frequencies. For example, photographs of a natural scene can be defined as a complex image. One of the earliest models that were developed to predict the degree of perceptual colour difference in images of complex scenes is the S-CIELAB /3/. The extension to complex images was performed by using a contrast sensitivity function (CSF). Hong and Luo /4/ developed a (Hue-angle) metric for complex images that assigns higher weight to dominant colours and to colours having a greater difference. Chou and Liu /5/ proposed a (P-CIELAB) metric for complex images that incorporates a visibility threshold for colour differences. This pixel-wise visibility threshold varies as function of chroma, local luminance gradient, and background uniformity. Fairchild and Johnson /6/ have presented probably the most advanced model for colour appearance of complex images. Their iCAM framework includes different sub modules accounting for various properties of images and viewing conditions in image analysis. The aim of this study was to test and compare the metrics or models for complex images in order to determine their capability to predict the degree of visually perceived colour differences in natural photographs. To the best of our knowledge, the study of Hardeberg et al. /7/ is the only published work where different state-of-the-art colour difference metrics of complex images have been compared to each other. They analysed the relation of CIELAB dE, SCIELAB, iCAM, Structural Similarity Index /8/, Universal Image Quality /9/ and Hue-angle metric /4/ with data from a psychophysical experiment in which the perceptual image difference was evaluated. They used six test images, but only two of these images were natural photographs. The rest of the images were more or less studio photographs or graphical images. Their results indicated that perceptual image differences cannot be directly related to colour image differences as calculated using the current metrics. We evaluate the state-of-the-art metrics narrowing the problem from that defined by Hardeberg et al. /7/. A known fact is that image content exerts an influence on image assessment. For example, portrait and landscape are typical views in natural photography. We selected only landscape type images for our study, as they satisfied our needs for the requirement of colour distributions and spatial contents. We wanted that colour distribution of the images is wide enough because colour distortions were made using different ICC colour profiles. We also wanted that the spatial content of the images covers a wide range because we wanted to test how the methods take into account the spatial details of the image. In addition, our psychophysical experiment tested the perceptual colour difference, nor the perceptual image difference. Implementation of the metrics The metrics that were investigated and compared in this study are listed in Table 1. Selected metrics can be divided into different classes based on differences and similarities in their functional properties. The standardized metrics that are based on CIELAB dE colour difference were not originally developed to address differences in complex images, but they were selected to form a baseline for the comparisons. These metrics include the CIELAB dE, CIE94 and CIEDE2000 metric /10/. 510 ©2010 Society for Imaging Science and Technology Table 1. The metrics that were used in the study Metric Intended use Reference dE Colour patches /10/ CIE94 Colour patches /10/ CIEDE2000 Colour patches /1/ Hue-angle Complex images /4/ P-CIELAB Complex images /5/ S-CIELAB Complex images /3/ iCAM Complex images /6/ In addition, the implemented metrics includes also CIELAB based metrics that were developed to predict the appearance of complex images. These are the Hue-angle, PCIELAB and S-CIELAB metrics. The first two are both similar in that they use a weighting scheme to address the structural properties of images, such that, a pixel-wise weight is applied to re-adjust a CIELAB dE value of the pixel to contribute to a more precise estimate of the perceived colour difference. But, as the Hue-angle metric computes the weight more globally, the P-CIELAB metric uses local properties of the image. The SCIELAB, which is also called a spatial extension to CIELAB colour space, takes advantage of the filtering characteristics of the human visual system (HVS) to apply the CIELAB dE metric to complex images. These characteristics are modelled with the contrast sensitivity function (CSF) in the frequency domain. Similarly, the iCAM framework uses the CSF, but instead of using the CIELAB colour space, it uses the IPT colour space. In addition, the iCAM framework consists of multiple modules that account for the viewing conditions and colour appearance phenomena. These include modelling of the chromatic adaptation, Hunt effect, Stevens effect, surround effect and lightness contrast effect. Test Images The distorted test images were created by changing their colours through ICC profiles gamut mapping process. Here, the absolute colorimetric rendering intent was used with four standard ICC profiles: Euroscale Uncoated, ISO Uncoated, PSR Gravure LWC, and Uncoated FOGRA. The gamut of these ICC profiles and the gamut of sRGB space in ab-plane are illustrated in Figure 1, where the visualizations have been obtained from ColorSync Utility included in Mac OS X. As can be seen from the figure, the ICC profiles can be divided into two groups based on their dimensions. This was done to ensure that there would be both larger and smaller differences between generated distortions. The selection of images for pilot tests from the candidate images was done by calculating average CIELAB dE values and then selecting those image sets that had average colour difference values on both sides of threshold value. The threshold value for colour difference discrimination in natural images is about 2.2 dEab /11/. Finally, the selection was further narrowed to eight images. Seven of the images were from the Photos.com –database /12/ and one was from the CIE TC8-03 test image set /13/. The test images are presented in Figure 2, where the images are named as Autumn road, Red field, Mountains, Forest rise, Red brushwood, Park, Table, and Picnic. (a) (b)