Color constancy refers to the capability to perceive the same color of an object under varying illumination conditions. Color constancy is a distinct feature in our color perceptual system which allows us to use color as a reliable feature of object for recognition. Double opponent neurons, found in striate cortex, are believed to play an important role in the color constancy neural mechanism. Its spectral and spatial opponent structure suggests that abrupt color variation, such as object' boundaries, measured in the color opponent space is the key to achieve color constancy....[ Read more ]

Color constancy refers to the capability to perceive the same color of an object under varying illumination conditions. Color constancy is a distinct feature in our color perceptual system which allows us to use color as a reliable feature of object for recognition. Double opponent neurons, found in striate cortex, are believed to play an important role in the color constancy neural mechanism. Its spectral and spatial opponent structure suggests that abrupt color variation, such as object' boundaries, measured in the color opponent space is the key to achieve color constancy.

Although the role of spatial opponency has been explained by retinex theory, the contribution of spectral opponency is still unknown. Using color edge representation, we have compared the double opponent neurons inspired model with the retinex model using an evaluation framework proposed. This evaluation framework not only assesses the extent of response constancy under varying illumination conditions, the discriminative quality among different color edges is also taken into consideration. The results indicate that incorporating spectral opponency with spatial opponency can give a better performance.

Next, we have developed a modified gray world model based on the double opponent neurons. We use the double opponent responses as an edginess measure which prevents overweighting the color of large objects. The performance of this model is compared against another three models, Moore's retinex, gray world and gray world RGB. We conclude that our model can discount illumination effect while maintain a good separation among different hues. The influence to performance by object's size is also reduced.