Predicting brightness of simple self-luminous scenes using an image colour appearance model

Phung Thanh Hang – KU Leuven

To overcome the challenges of applying object colour appearance models (CAM) on self-luminous stimuli such as the ambiguity of the definition of the reference white or the underestimation of the Helmholtz-Kohlrausch effect, several colour appearance models dedicated to self-luminous stimuli have been developed. Yet, these models only consider a simple test scene, while the self-luminous scenes are much more complex. Therefore, Image Colour Appearance Models have been introduced to predict the colour appearances of complex images for various imaging applications. Among these, iCAM is a model which can output, pixel by pixel, the same perceptual attributes as traditional CAMs (brightness, hue, colourfulness, lightness, chroma and saturation). This study presents an evaluation of iCAM’s performance when applied to a simple self-luminous scene, in predicting the influence of background luminance, background size, stimulus saturation, and stimulus size on stimulus brightness.
A set of virtual relative XYZ images representing a scene of 5x3m, in which 1 pixel corresponds to 1x1cm in the real setup, is created as the input for the model evaluation. The relative XYZ values of the real stimulus are associated to those pixels corresponding to the stimulus and the same principle is applied to the background. The first test is performed on 90 scenes to predict the brightness of neutral stimuli presented on neutral backgrounds with different stimulus and background luminance values. Different filter kernel sizes are also used in this test. The predictions of iCAM for other phenomena are also reviewed in the experiments 2-4. 30 coloured stimuli from 6 hues are used to test the prediction of the Helmholtz-Kohlrausch effect, while the background size effect evaluation is performed with multiple background sizes (0%, 12.5%, 25%, 50%, 75% or 100% of the image size), and the stimulus size effect prediction is tested on 40 scenes with various stimulus sizes (1˚, 2˚, 3˚, 5˚, 10˚, 15˚, 20˚, 25˚ and 30˚).
The results indicate that iCAM predicts the background luminance and background size effect well, with the exception of unrelated stimuli (dark background). A quick investigation on the effect of filter kernel size reveals that the choice of filter kernel size seems to be linked to the physiological mechanism of image processing of the visual system. Further study is needed to justify this argument. The model is also capable of predicting the background size effect since the brightness of the stimulus is predicted to decrease when a bright background increases in size. Although the Helmholtz-Kohlrausch effect is not explicitly included, the model can predict the effect successfully for blue and cyan, but not for other hues. Furthermore, the impact of stimulus size on brightness is underestimated as no change is observed when the stimulus size increases. Hence, this calls for a new image colour appearance model dedicated to complex self-luminous scenes.