Learning image aesthetic subjectivity from attribute-aware relational reasoning network

摘要

People's various visual preferences lead to subjectivity in their aesthetic perception of images. Hence, learning image aesthetic subjectivity has attracted great interest in the computer vision community. People with different subjectivity often have huge uncertainty in their aesthetic ratings, which is affected by diversified aesthetic attributes in images. Although existing studies leverage aesthetic attributes to infer image aesthetic scores, the influence mechanism of diversified aesthetic attributes on people's aesthetic ratings has not yet been revealed. Because of this, this paper proposes an attribute-aware relational reasoning network to learn image aesthetic distribution rated by people with different subjectivity. Specifically, we embed the relationship between different aesthetic attributes into a deep network for reasoning the uncertainty in image aesthetic distribution. Besides, an efficient distribution loss function is introduced to intensively learn image aesthetics with high uncertainty. Experimental results show that our method is superior to state-of-the-art methods in learning image aesthetic distribution. (C) 2022 Elsevier B.V. All rights reserved.