Technical Quality-Assisted Image Aesthetics Quality Assessment

Image aesthetics assessment (IAA) aims at predicting the perceived aesthetic quality of images. Intuitively, the technical quality of an image has significant impact on its aesthetic quality, e.g., an image with noticeable distortions is not likely to have very high aesthetic quality. However, this characteristic has rarely been considered when designing modern IAA models. Motivated by this, this paper presents a new Technical Quality-assisted multi-task deep network for image Aesthetic Quality assessment, dubbed TQ4AQ. Specifically, we first extract theme-aware general aesthetic features based on the attention mechanism. Meantime, hand-crafted technical quality features are extracted from aesthetic images. Then the general aesthetic features are utilized to predict the technical quality features and the aesthetic quality simultaneously, based on which technical quality features are integrated. By this means, the aesthetic features are empowered the capability of understanding technical distortions, and more comprehensive aesthetic feature representations are obtained for IAA. Extensive experiments demonstrate the advantage of the proposed TQ4AQ model over the state-of-the-arts.