Abstract
Correctly interpreting retinal images requires distinguishing between various types of contours, including surface boundaries, within-surface reflectance changes, and shadows. There are many hypotheses about how the visual system does this, but few attempts to directly measure the relevant information in natural images. Thus, we have analyzed close-up images of foliage obtained with a calibrated 36-bit camera, which estimates relative L, M, and S cone responses with error SDs of 0.15%, 0.1% and 1%, respectively, for natural spectra. Cone responses were transformed into a logarithmic space having orthogonal dimensions: luminance, blue-yellow, and red-green (Ruderman et al. JOSA A, 15, 2036). More than 2000 leaves were hand segmented from over 60 images representing a wide range of foliage; then, within each leaf, reflectance contours and shadow contours were hand traced. To compare the different types of contour, we extracted a ribbon of pixels a few pixels wide on each side of each contour. For each color channel, and various ribbon lengths, we measured the difference in the mean, rms amplitude, and spatial coherence across each contour. We found larger rms amplitude and coherence differences across surface boundaries than across surface-reflectance or shadow contours, and larger mean differences across shadow contours and surface boundaries than across surface reflectance contours. The ribbon information was quantified by the signal-to-noise ratio (d') for deciding whether or not a ribbon was of one type as opposed to one of the other two types. The d' values were substantial for all three types of contour.
Supported by NIH grant EY11247.