Abstract

According to the adaptation theory, changes in the sensitivity of the cones caused by illumination playa very critical role in color constancy. However, since the sensitivity of the cones may rapidly dissipate as the levels of adaptation grows up, adaptation theory assumes that eye-movement is an essential part controlling the sensitivity of the cones not to be saturated. A neural network model for color vision was developed to simulate the function of adaptation and eye-movement as specified in adaptation theory. Color constancy was the focal interest in the designing of the neural network A six-layered neural network was designed in such a way to accomodate a number of human principles of color proce:3sing from retina to cortex. Units in each of the six layers were made to mimic the three kinds of cones, adaptation of cones, color-opponent units, double-opponent units, interconnecting units between area 17 and V4, and the units in area V4, respectively. The fifth layer were treated as hidden and its connections with the fourth and sixth layers were made to be adaptively determined by back-propagation algorithm, Outputs of the network were mapped onto the CIE color space by the units in the sixth layer. The effect of eye-movement was simulated by shifting the input color image back and forth across the input field. Hence, the neural network thus developed was dynamic in nature. In order to evaluate the effect of installing adaptation and eye-movement mechanisms in the neural network on approximating color constancy, the performance of the dynamic neural network was compared with that of a static one of which structure is identical to the dynamic one except the absence of adaptation and eye-movement mechanisms. The performance of the dynamic neural networks approached to color constancy in a satisfactory level and was significantly superior to that of the static neural network.