Predicting emotion ratings from color statistics of images

We utilized this data to see how visual elements affect our everyday lives by engineering features of images and running that through a variety of neural networks.

Which features in art affect emotional reactions? In order to answer this question, we took a variety of abstract images and included people’s emotional ratings for each image, for eight different emotions. We utilized this data to see how visual elements affect our everyday lives by engineering features of images and running that through a variety of neural networks. Out of all the networks we ran our data through, the linear regression model performed the best with a mean average error of around 1.4. All the other neural networks, including a convolutional neural network, had a mean average error of at least 1.7. The full-size linear regression model with all engineered features worked best in predicting the emotional values associated with each image. That is, all specific features were significant and we did not find meaningful differences among the models for different emotion ratings.

Explore More!

Research Paper