Fig 1.
a) Examples of RGB natural color images from the database. b) Set of N = 50 filters obtained after digitizing to 1 bit the luminance (L+M) coordinate. The actual bandwidth occupancy of the set () turns out to be slightly higher (0.06) than the imposed limit W (0.05). c) Set of N = 50 filters obtained after digitizing to one bit the l (L/(L+M)) coordinate. The actual bandwidth occupancy of the set out to be slightly higher (0.06) than the imposed limit W (0.05), but is still the same as the actual bandwidth of the 1 bit luminance set. d) Two gray-levels luminance sketches obtained with the filters in (b). e) Color-only sketches obtained with the filters in (c). f) Set of N = 50 filters obtained after digitizing to 2 bit the luminance coordinate. The actual bandwidth occupancy of the set turns out to be slightly higher (0.07) than the imposed limit W (0.05). g) Set of N = 50 filters obtained after digitizing to 1 bit both l (L/(L+M)) and luminance (L+M) coordinates. The actual bandwidth occupancy of the set turns out to be slightly higher (0.07) than the imposed limit W (0.05), but is still the same as the actual bandwidth of the 2 bits luminance set. h) Four gray-levels luminance sketches obtained the set of filters in (f). i) Color sketches obtained with the set of filters in (g).
Fig 2.
Information content of 1 bit sketches.
Upper panel: distributions of the information content of the 1 bit color (orange) and luminance (gray) sketches. Information of a single sketch, S(i, j), is calculated as , where k (i,j) is the pattern matching the patch centered at pixel (i,j), summed over all the image. The distributions are taken over the entire image database of our study. Lower panel: percentage of correct discrimination, averaged over all subjects, plotted as a function of the information content of the color only (orange) and luminance (gray) sketches, for the same data as in Fig 3. Error bars are SEM. The dashed line represents chance level.
Fig 3.
Discrimination of images based on luminance and color sketches.
Percentage of correct discrimination of three subjects in four different conditions tested: 1bit luminance-only sketches (2 gray-levels bars), 1 bit equiluminant sketches (2 red/green-levels bars), 2 bit luminance-only sketches (4 gray-levels bars) and 1 bit equi-luminance + 1 bit luminance sketches (4 red/green-levels bars). Values correspond to averages over different sessions (300 trials) of the same condition. Errors are binomial s.d. The dashed line represents chance performance. Statistical binomial tests. 1bit luminance vs. 1bit equiluminance: p<0.0001 for each subject; equiluminance vs. chance p>0.1; 1 bit luminance vs. 1 bit equiluminance +1 bit luminance: N.T and L.S p>0.05, M.D p = 0.048;1 bit luminance vs. 2 bits luminance: and L.S p<0.0001, N.T and M.D p>0.1.
Fig 4.
a) Luminance features vs. color features rendered in luminance. Percentage of correct discrimination of five subjects based on 2 gray-level sketches obtained with 1 bit luminance-only filters (Fig 1b) (black-outlined bars) and 1 bit color-only filters (Fig 1c) rendered in gray-levels (red outlined bars). Plotted values are averages over all trials (300) of the same condition. Errors are binomial s.d. The dashed line represents chance performance. Statistical binomial tests: N.T. and G.C: p< 0.05, D.B., M.L. and M.D. p<0.01. b) Color features vs. color patches. Percentage of correct image discrimination based on equiluminant sketches (2 red/green-levels bars) and percentage correct discrimination of color patches (blue bars). Data collected on three subjects. Plotted values are averages over all trials (300) of the same condition. Errors are binomial s.d. The dashed line represents chance performance. Color patches vs. chance: binomial tests p<0.005 all subjects.
Fig 5.
Discrimination of sketches obtained with optimal patterns (left) and of sketches where optimal patterns were replaced by randomly chosen non-optimal patterns (right). Each color represents a different subject.
Fig 6.
The gray histogram represents the probability distribution, in natural log scale, of the 227 possible 3x3 patterns, extracted from the dataset of RGB images used in this study [16]. The green curve is the model selection function with N = 500, W = 0.08. The green histogram is the probability distribution of corresponding selected patterns. The red histogram is the probability distribution of equiluminant patterns.