Aesthetic evaluation network with refined style attributes for art images

Abstract

Most deep learning-based aesthetic image evaluation (AIAA) methods for art images fail to effectively extract style attributes from art images, as existing art image datasets rarely provide style-attribute annotations. The lack of style attribute information for art images reduces the interpretability of AIAA methods and limits their performance. To address these issues, this paper proposes an aesthetic evaluation method for art images based on style attributes. The method comprises a feature extraction module, a feature weight fusion module, and a joint prediction module. The feature extraction module is pre-trained on both general aesthetic image datasets and art image datasets to learn style attributes and general aesthetic information. To more effectively fuse the extracted features, the feature weight fusion module adaptively assigns weights to emphasize significant style-related information. The joint prediction module employs a nonlocal self-attention block to fuse style attribute features with general artistic aesthetic features. The fused features are fed into a multilayer perceptron (MLP) layer to compute the aesthetic score of an art image. Experimental results demonstrate that the proposed method outperforms several state-of-the-art AIAA approaches.