Predicting Scores of Various Aesthetic Attribute Sets by Learning from Overall Score Labels

For aesthetic attribute evaluation of images (AAEI), the annotation of image aesthetic attribute scores plays an important role. It requires experienced artists and professional photographers, which hinders the collection of large-scale fully-annotated datasets. In this paper, we propose to learn an AAEI model from the datasets with only the overall aesthetic scores of the whole image. First, a novel aesthetic attribute evaluation framework based on attribute features is proposed to predict attribute scores and overall scores. We call it the F2S (attribute features to attribute scores) model. Our F2S model is trained from instances with only overall aesthetic scores, bridging the gap between overall and attribute scores. Second, we define an aesthetic attribute contribution to describe the role of aesthetic attributes throughout an image and use the attention weights from a channel attention module as contribution values. Last, we propose the relationship among the attribute scores, the attribute contributions, and the overall scores and use the relationship to train our F2S model. Sufficient experiments on publicly available datasets demonstrate that our F2S model achieves comparable performance with those trained on the datasets with fully annotated aesthetic attribute score labels. Using our method, it is feasible to learn meaningful attribute scores for various aesthetic attribute sets of different types of images with only overall aesthetic scores, such as subject and composition for landscapes and emotion and pose for portraits.