Augmenting Image Aesthetic Assessment with Diverse Deep Features

With the increasing prevalence of digital images, automatically assessing the aesthetic quality of photos could benefit many real-world applications. While many previous methods have produced binary classification results, this paper proposes a model to produce regression results with high accuracy. The proposed model exploits global visual information such as color palette, saturation, and clarity, as well as deep features like blur maps, saliency maps, and scene information to augment the DenseNet architecture. The augmented DenseNet, when evaluated on the AVA dataset, outperformed the current state-of-the-art methods, achieving an accuracy of 88.65% on the 10% subset and a Spearman's rank correlation coefficient of 0.5802 on the full dataset. Comparison of the augmented DenseNet and the DenseNet baseline also demonstrate the effectiveness of the proposed methods of augmentation.