Predicting Human Visual Complexity Judgments via Deep Learning

Visual complexity can be defined as the difficulty of defining an image or the number of details in the image. Studying visual complexity is important for understanding human visual system and developing systems for human use. In this study, a deep learning model developed for object identification was used in the context of transfer learning for developing two models to predict human visual complexity judgments. The model that was trained for making prediction within the same domain is named as within-domain, while the model that was trained for making prediction between different domains is named as cross-domain model. After the training phase, within-domain model can predict which image is more complex than the other with 95% accuracy and the cross-domain model with 78% accuracy. Visual complexity scores of images that are not in the test set were calculated by means of image complexity comparisons of models. It was found that the correlation between real visual complexity scores and calculated scores was strong, even for object (0.77) and scene (0.83) categories.