Analysis of Deep Features for Image Aesthetic Assessment

The performance of image aesthetic assessment has been significantly improved by deep learning techniques in comparison to traditional hand-crafted feature-based methods. However, there has not been an attempt to analyze the learned features in deep learning approaches. This paper presents in-depth analysis of the deep models and the learned features by the models for image aesthetic assessment in various viewpoints. We consider binary classification of the mean (average aesthetic level) and standard deviation (subjectivity of aesthetic perception) of aesthetic ratings. In particular, our analysis is based on transfer learning among image classification and aesthetic classifications for comparative analysis. Our results highlight the similarity and transferability of learned features among the classification tasks, and comparison of the trained models and convolutional filters.