Fig 1.
A biological plausible computational framework for color perception.
The model initiates with simulating the responses of single and double opponent cells to a visual stimulus in chromatic (red-green and blue-yellow) and achromatic channels. The DO and the intensity channels are reconstructed using SNNs and linearly combined with the SO channels to provide the perceived image. Image by Alexander Ivanov (Pixabay).
Fig 2.
V1 recording and simulation results during perceptual filling-in of black and red squared surfaces.
A. Averaged early (60–100 ms following stimulus onset) Macaque V1 VSDI-measured neural activity map following exposure to black (left) and red (right) squared surfaces with various sizes (0.5°– 8°) (see Methods); B. Spatial profiles crossing through the edges and center of the V1 activation patches. The continuous vertical line marks the peak activation position in the 0.5° square response profile, which corresponds to the center of square in larger stimuli. Responses to a 2° square are marked with vertical dashed lines; C. Model-derived results for the reconstruction of black and red squared surfaces with various sizes squares; D. Cross-sectional profiles of each reconstructed square along the x-axis.
Fig 3.
Visual stimuli are presented in the first row and the reconstructed images are shown below, generated with various values of αc, βc, αi and βi (aligned to the left of the corresponding images). LPIPS scores are shown on the right.
Fig 4.
Reconstruction of the cube illusion.
A. The original cube images under three illuminations: natural, yellow, and blue (First row). The model predictions with different sets of chromatic and achromatic parameters are shown in rows 2–5; B. Comparison between the true color (marked with an asterisk) and the predictions of the model with αc = 1 and αc = 0.7. Results are presented in u’v’ (CIELu’v’) color space. Each color circle surrounds the true and predicted colors of a sampled pixel in the patch. Black lines represent cone-opponent axes, S/(L + M) and L/(L + M). The intersection of the lines represents the achromatic point.
Fig 5.
Reconstruction of the color assimilation grid illusion.
A selective colored grid was overlaid on two grayscale images, creating the visual stimuli, shown in the first row. The model predictions with different sets of chromatic and achromatic parameters are shown in rows 2–5. LPIPS scores are shown on the right.
Fig 6.
Reconstruction of #TheDress and #theShoe photos.
A. The original images; B. Model’s prediction with different sets of chromatic and achromatic parameters; C. Comparison between the true color (marked with an asterisk) and the predictions of the model Results are presented in u’v’ (CIELu’v’) color space. Each color circle surrounds the true and predicted colors of a sampled pixel in the patch. Black lines represent cone-opponent axes, S/(L + M) and L/(L + M). The intersection of the lines represents the achromatic point. b) Comparison between the true color (mark with red *) and the predictions of the model with αc = 1 and αc = 0.5 presented in u’v’ (CIELu’v’ 1976) color space. Each ellipse surrounds the true and the predicted colors of a sampled pixel in the patch. Black lines represent cone-opponent axes, S/(L + M) and L/(L + M). The intersections of the lines represent the achromatic point.
Fig 7.
Parameter evaluation with 21x21 and 11x11 kernel size (2nd and 3rd columns, respectively), compared with Retinex prediction (right column).
Input images are shown in the left column.