Higher Order Color Mechanisms for Image Segmentation

Abstract

Recent physiological evidence has shown that neurons at the early visual stages are selective for a combination of color, luminance and orientation. Neurons with a linear response tuning, resulting in broad tuning curves, are found at all stages, but the proportion of nonlinear neurons, narrowly tuned for color, increases along the visual pathway. We ran psychophysical experiments to characterize the number and tuning widths of the mechanisms underlying image segmentation. We used a noise masking paradigm with different types of noise to disentangle mechanisms with narrow and broad tuning characteristics. The data were best described by a chromatic detection model with multiple, broadly tuned mechanisms, where narrow tuning curves emerge due to off-axis looking. We then analyzed a set of calibrated natural images and determined the joint statistics of color and luminance edges. The majority of edges in natural scenes was characterized by a contrast in both color and luminance, while some prominent object boundaries were signalled only in the chromatic plane. Based on the converging evidence from different disciplines we conclude that multiple linear, broadly tuned mechanisms which are selective for a combination of chromatic contrast, luminance contrasts and orientations play a central role for contour extraction and robust image segmentation.