Representation learning of image composition for aesthetic prediction

Photo quality assessment (PQA) aims at computationally and precisely evaluating the quality of images from the aspect of aesthetic. Image aesthetic is strongly correlated with composition. However, few existing works have taken composition into consideration. Besides, existing composition features are typically hand-crafted. In this paper, we propose a novel end-to-end framework for representation learning of image composition. Specially, we build a fully connected graph based on deep features in Convolutional Neural Networks (CNNs). In the graph, edge attributes i.e. similarities between deep features at different positions are used for representing image composition. Besides, we use global attributes of the graph to represent miscellaneous aesthetic aspects. Finally, we use a gate unit to combine both composition features and miscellaneous aesthetic features for aesthetic prediction. The whole network can be trained in an end-to-end manner. Experimental results show that the proposed techniques significantly improves the prediction precision of aesthetic and composition over various datasets. We have released our codes at: .