2012 年 11 巻 1 号 p. 103-111
In this study, we focus on the sensitivity characteristic of the L-, M- and S-cones in human color vision, and construct a model that outputs two sets of Kansei values characterized as “bright - dark” and “light - heavy” using a four-layered neural network. In a subjective test, examinees evaluated 64 color patterns on a CRT display using two word pairs on a seven step rating scale. The evaluation results were standardized (min. -0.9, max. 0.9) as the output of the neural network. Three units of the input layer in the network model corresponded to the L-, M- and S-cones, respectively. The values of the absorption spectra characteristics of the L-, M- and S-cones were fed into each input unit. After training, the output of the model for both training results and performance test results were very close to the teaching Kansei values. We confirmed that the model has reproduced human Kansei values for colors. Furthermore, analyzing the synaptic weights values of the neural network model after training, we obtained the units that extracted individual differences of two color experts.