Multi-Task Learning of Visual Attributes for Image Aesthetics Assessment

The complexity of individual aesthetic preferences for images is still not fully understood, making image aesthetic assessment a difficult challenge in AI. While many factors contribute to these preferences, certain aesthetic attributes play a key role. We also found that image composition significantly affects these preferences. Composition involves arranging different parts of an image, such as regions or objects, to create harmony and unity. In this research, we present a multi-task convolutional neural network that integrates these aesthetic attributes. This approach promotes better generalization by using shared representations. We propose a fully connected adaptor based on deep features to represent image composition, enabling end-to-end training with Pairwise Ranking loss and Euclidean loss. Additionally, we use a gate linear unit to merge composition and other aesthetic features for prediction. Our experiments show that the proposed model achieves the highest Spearman's rank correlation coefficient (SRCC) and linear correlation coefficient (LCC) for both overall aesthetic score and attribute score predictions across multiple datasets.Moreover, the multi-task network consistently outperforms in SRCC and LCC, highlighting the benefit of learning multiple aesthetic attributes simultaneously.