Detecting color in natural scenes
Résumé
The human visual system processes both luminance and chrominance information from our environment. Three types of retinal photoreceptors (L, M and S cones) convert light into three channels: a luminance and two chrominance channels through cone-opponent mechanisms. These channels are further processed by the cells of the primary visual cortex. Several psychophysical studies [1,2] have modeled these post-receptoral channels by linear combinations of cone-contrast values. However, when considering natural scenes, cone-contrast definitions appear to be arguably unsuitable, mainly because adaptations levels of cones are likely to vary across the image. To perform pixel-based analyses on luminance and chrominance information in natural scenes, some studies [3,4] referred to a non-linear "shadow-removing" definition of chrominance proposed by Párraga et al. [5] and included a divisive normalization of the color-opponent channels by luminance. In the proposed study, we access the relevance of this divisive model to the perception of natural scenes. To tackle this issue, we introduced three experiments where we manipulated luminance and chrominance channels contents of natural scene images and studied their relationship with perceived information. In a first experiment, using a 2AFC paradigm, subjects were asked to decide which of the two displayed images of the same scene contained color (each image having the same luminance information). Our data show that color detection threshold increases with the mean luminance level of the image. To account for these data, we suggest a simple but effective model for color detection in our task. Our model has two variants: one is based on a linear color space, while the other features divisive normalization of chrominance by luminance (along the lines of Párraga et al.). We found an advantage for the divisive definition, but further analysis showed that a misalignment of individual cone-opponent channels with our color space cardinal directions could also account for our results. To discard this explanation, we conducted two additional experiments of luminance and chrominance detection in natural scenes. Our findings suggest that divisive definitions of color-opponent channels are indeed more appropriate than linear definitions when considering performance on detecting color in natural scenes. However, another model could also account for our data: if each cone logarithmic signal independently follows a von Kries adaptation procedure over the image, it produces results very similar to a divisive model, raising the question of which mechanistic description of cortical color vision is the most relevant.