Abstract
Few studies have investigated the structural relationships between modeled neural images of the luminance, red-green and blue-yellow post-receptoral channels in response to natural scenes. Here we examine these relationships for both first-order, i.e. luminance and color, and second-order, i.e. texture and contrast, variations in a set of natural color images.
Images collected using a calibrated digital camera were transformed into LMS cone responses for each pixel, which were then converted into luminance, red-green, and blue-yellow channel images. Simulated responses of cortical first- and second-order operators were produced by convolution with linear filters (Gabor functions) or filter-rectify-filter operators, respectively, for a wide range of filter orientations and spatial frequencies. Filter response amplitudes and image statistics (kurtosis and entropy) were examined, as well as ‘signed’ and ‘unsigned’ cross-correlations between the three first-order channel images and between the first- and second-order channel images.
The results demonstrate that first-order red-green has a higher kurtosis/entropy than blue-yellow, which in turn has higher values than luminance. Correlations between first-order luminance and first-order color information are surprisingly high. Additionally, first-order luminance and color are strongly correlated with second-order luminance, but not second-order color. These results suggest that higher-order chromatic statistics play a distinct role in natural images.
Funded by NSERC grant to CLB (OPG0001978) and a CIHR grant to FAAK (MOP-11554).