PEAR: Photographic Embedding for Aesthetic Rating

Image aesthetic quality assessment has gained the enormous interest in recent years with the advancement of deep learning and the large-scale datasets. Current state-of-the-art methods generally leverage deep features, while the hand-crafted photographic attributes which are also useful but not always available have not drawn the sufficient attention. In this paper, we propose the Photographic Embedding for Aesthetic Rating (PEAR) framework to assimilate their advantages. In PEAR, a prior network is constructed to smoothly inject the photographic attributes to the latent space, and handle the attribute missing case via knowledge transfer. Simultaneously, aesthetic quality assessment is formulated as multi-task learning of aesthetic rating and image reconstruction to regularize the learning of latent codes. The extensive experiments on two challenging datasets demonstrate PEAR outperforms the state-of-the-art methods.