Modeling Image Composition for Visual Aesthetic Assessment

Composition information is an important cue to characterize the aesthetic property of an image. We propose to model the image composition information as the mutual dependencies of its local regions, and design an architecture to leverage such information to boost aesthetics assessment. We adopt a Fully Convolutional Network (FCN) as the feature encoder of the input image and use the encoded feature map to represent the individual local regions and their spatial layout in the image. Then we build a region composition graph in which each node denotes one region and any two nodes are connected by an edge weighted by the similarity of the region features. We perform reasoning on this graph via graph convolution, in which the activation of each node is determined by its highly correlated neighbors. Our method achieves the state-of-the-art performance on the benchmark visual aesthetic dataset [15].