Color Transparency

Author

Peggy Gerardin

Introduction

The visual system identifies surfaces and uses their properties to help recognize objects. One example of this phenomenon is color transparency: when a surface is seen both in plain view and through a transparent overlay, the visual system still identifies it as a single surface. Several studies have suggested that color changes across a region of an image that can be described as translations and/or convergences in a linear trichromatic color space lead to the perception of transparency, but other transformations, such as shear and rotation, do not. To study the limits of such systemic chromatic changes, we generated classes of stimuli consistent and inconsistent with D'Zmura et al's Convergence Model of transparency perception. The main results support the Convergence Model in showing that, for vectors exceeding a minimal length, convergence and translation (except when equiluminant) lead to the perception of transparency, while shear and divergence do not. Large equiluminant translations were less often judged as transparent, consistent with observations reported by Chen and D'Zmura with respect to color changes that cross hue boundaries. Surprisingly, we found that small shears and divergences were also classified as transparent, in contradiction with the model. This could imply that two mechanisms underlie the perception of transparency: a low contrast mechanism that is sensitive to chromatic and luminance change independent of its direction (translation, shear, convergence or divergence), and one sensitive to higher contrasts that depends on the integrated direction. Our goal is to establish a model of color transparency, defining optimal conditions where this phenomenon occurs, and to investigate its application to color image rendering algorithms.

Related work

Transparency and motion
Transparency and shadow

Main contributions

This research will be useful to solve problems in color gamut transfers between devices (gamut mapping) and image segmentation. Image segmentation is the first step in image analysis and pattern recognition. The actual methods are based on histogram thesholding, region or edge-based approaches, and then image-independant. With our research, we will try to find an image-dependant color segmentation method that allows separating illuminance from surface colors, and mapping individual image components separately.

Current and future research

Following is a short summary of the next steps:

• Extension of this model to moving color transparency and its consequences/implications to color changes (psychophysical experiments and data analysis). Some studies have added motion in their stimuli, claiming that this enhances the transparency effect. The main goal of this experiment 3 is to expect whether motion is neutral with respect to the effects of systematic color changes. At the same time, a direct application of the psychophysical results on moving color transparency detection will be processed. The goal of this application is to test the robustness of our D'Zmura Model extension to real images and video sequences.
• Extension of this model to shadow perception, variation of size, form and uniformity (psychophysical experiments and data analysis). The main goal of this experiment 4 is to define the close relation between transparency and shadow perception. Both seems to need the same constraints, as the configuration (X-junctions) and systemic color changes. In this way, a new approach to identify and classify shadows in digital images will be processed, in using a perceptual model. The goal of this application is also to test the robustness of our D'Zmura Model extension to real images to detect shadows.

Collaborations

Sabine Süsstrunk
Ken Knoblauch (INSERM, France)

Publications

Funding

Swiss National Science Foundation (SNF) under grant number 20-59038.99.