Theme-Aware Semi-Supervised Image Aesthetic Quality Assessment

Image aesthetic quality assessment (IAQA) has aroused considerable interest in recent years and is widely used in various applications, such as image retrieval, album management, chat robot and social media. However, existing methods need an excessive amount of labeled data to train the model. Collecting the enormous quantity of human scored training data is not always feasible due to a number of factors, such as the expensiveness of the labeling process and the difficulty in correctly classifying data. Previous studies have evaluated the aesthetic of a photo based only on image features, but have ignored the criterion bias associated with the themes. In this work, we present a new theme-aware semi-supervised image quality assessment method to address these difficulties. Specifically, the proposed method consists of two steps: a representation learning step and a label propagation step. In the representation learning step, we propose a robust theme-aware attention network (TAAN) to cope with the theme criterion bias problem. In the label propagation step, we use preliminary trained TAAN by step one to extract features and utilize the label propagation with a cumulative confidence (LPCC) algorithm to assign pseudo-labels to the unlabeled data. This enables use of both labeled and unlabeled data to train the TAAN model. To the best of our knowledge, this is the first time that a semi-supervised learning method to address image aesthetic assessment problems has been studied. We evaluate our approach on three benchmark datasets and show that it can achieve almost the same performance as a fully supervised learning method for a small number of samples. Furthermore, we show that our semi-supervised approach is robust to using varying quantities of labeled data.

Automated summary

Image aesthetic quality assessment is widely used in various applications, but existing methods require excessive labeled data. This study proposes a theme-aware semi-supervised method to address this issue, achieving comparable performance with fully supervised learning. This method consists of a representation learning step and a label propagation step.