Abstract
Purpose. We explored the statistical structure of human cone responses to natural images, particularly with regard to the analysis of space and color. Methods. Seven different temperate woodland scenes were imaged using a small-CCD camera in conjunction with a variable interference filter. Each data set contained 43 128xl28-pixel images taken at 7-8 nm intervals from 403 to 719 nm, with object reflectances within the scene standardized to a white diffuse reflector. Images were transformed into human L, M, and S responses using Stockman-MacLeod-Johnson fundamentals, converted to log 10 values, and scaled relative to the mean logtransformed value. These new values (L, M, S) were subjected to principal components analysis. Results. We found the principal components to correspond to a luminance signal (0.581L + 0.581M + 0.5705), a blue/yellow chromatic opponent (0.407L + 0.400M - 0.8215) and a red/green opponent (0.705£, - 0.709M + 0.004S), with variances in the ratio of 45:9:1. These components vary little with spatial scale, up to blocks of 16x16 pixels. Furthermore, when all 3x3 patches of images in our data set were subjected to analysis, each of the resulting 27 9-pixel blocks consisted only of luminance, blue/yellow, or red/green values, suggesting a complete separation of spatial and chromatic components. Conclusions. Intensity, blue-yellow opponent, and red-green opponent signals from human cones are not correlated with each other, and are independent of spatial scale. Furthermore, in small image patches chromatic and spatial signals can be decorrelated. Thus, primate opponent processing systems, which form luminance, blue/yellow, and red/green systems, are appropriate for efficient transmission and analysis of information contained in natural scenes. Supported by NSF Grant IBN-9413357.