Multi-theme image aesthetic assessment based on incremental learning

Image aesthetic assessment is a complex and subjective task. Image theme categories affect observers' aesthetic judgments during human visual perception. Considering the diversity of image themes, image aesthetics are correlated with themes. Learning aesthetics directly from images ignores the impact of thematic variation. We hope to obtain more comprehensive aesthetic features for all thematic images through differentiated learning by dividing images into themes and then constructing a complete aesthetic assessment model, which can be more reasonable for the aesthetic assessment of images. Therefore, we propose a multi-theme image aesthetics assessment method based on class-incremental learning (CILNet). The images are sub-thematically trained according to the theme during the training process to learn the aesthetic characteristics of the theme. The feature extraction network in each incremental step performs feature extraction and pruning based on the importance of the theme for each theme image. It saves the theme-related importance aesthetic feature vector. The model compiles comprehensive multi-theme aesthetic features as the theme images are progressively added. Finally, aesthetic scores are predicted based on these compiled multi-theme aesthetic features. We trained and tested CILNet on generic and multi-themed image datasets and compared its performance with other aesthetic assessment models. The experimental results prove that CILNet can effectively obtain the aesthetic features of each theme image and effectively improve the performance of the aesthetic assessment model. Compared with the state-of-the-art model, CILNet is lightweight and achieves similar performance.