Self-Supervised Image Aesthetic Assessment Inspired by Aesthetic Domain Knowledge

Abstract

Photographic rules describe how to create high-quality images by imposing restrictions on some aspects like lighting and color. Such rules referred to as domain knowledge, turn out to be crucial in enhancing the performance of image aesthetic assessment. Although many research efforts have been made, aesthetic domain knowledge is underutilized, and a large amount of labeled data are typically required. To remedy these issues, we propose an improved multi-task self-supervised method under the guidance of aesthetic domain knowledge. In the pre-training phase, we design multiple pretext tasks for a naïve network to predict the levels of the properties related to photographic rules. That enables the network to learn visual features sensitive to image aesthetic information. After that, the well-trained network is applied to evaluate the aesthetic quality of images. Experiments on two benchmark datasets elucidate that the features learned by the network can discern aesthetic images according to their difference in photographic properties. The promising results demonstrate that the proposed method can successfully leverage aesthetic domain knowledge to learn effective features from large-scale unlabeled data for image aesthetic assessment.