Abstract
Humans tend to group nearby image patches that are similar in color and segregate nearby image patches that are dissimilar in color. However, little is known about the predictive power of this chromatic information in natural scenes. Thus, we have analyzed close up images of foliage obtained with a calibrated 36-bit camera, which provides estimates of the relative L, M and S cone responses with error SDs of 0.15%, 0.1% and 1%, respectively, for natural spectra. More than 2000 leaves were hand segmented from over 60 images representing a wide range of foliage. Image patches were sampled from each leaf object. For each patch, other patches were sampled within and outside the leaf. The cone responses for each patch were transformed into a logarithmic space (Ruderman et al. JOSA A, 15, 2036). Probability distributions for response differences (based on all objects) were estimated, conditional on distance between patches and on whether or not patches fall within the same leaf. The grouping/segregation information was quantified for each patch size and distance by the signal-to-noise ratio (d') for deciding whether or not a pair of patches was drawn from the same leaf. For small patches, d' fell from values greater than 2.0 at smaller distances to values near 1.0 at larger distances. Thus, for foliage (a large fraction of the natural world) there is substantial chromatic information for region grouping. Further, the decline in d' with distance suggests that region grouping would benefit from a chromatic region-growing mechanism.
Supported by NIH grant EY11247.