Composition-aware Learning for Aesthetic Assessment of Edited Photos

It is common that people prefer to edit photos before posting, and the aesthetic assessment of edited photos is increasingly important, which has applications in photo recommendation, retrieval, and automatic beautification. To our knowledge, this is the first attempt to study the problem of aesthetic assessment for edited photos. Since no related public dataset is available, we create a large-scale Edited photo Aesthetic Data set (EAD) including 25,000 images by incorporating a diversity of personal judgments from the photo-sharing platform. Moreover, in this paper, we propose a new deep learning framework MSNet to obtain the composition characteristics of the edited photos in aesthetic assessment. Specifically, there are two main modules in MSNet. The global composition information describing the overall picture is extracted through the Multi-level feature Fusion Module while the Spatial Attention-aware Module is used to get the local composition information expressing the dependence between visual elements. Finally, the experimental results show the superiority of our proposed approach against the traditional methods designed for non-edited photos.