Spatial Attentive Image Aesthetic Assessment
Ying Xu, Yi Wang, Huaixuan Zhang, Yong Jiang · 2020
Image aesthetic assessment is challenging as it concerns various relative vague perceptual evaluations. A key factor among them is how to evaluate image layout, finding spatial importance in aesthetics. In this paper, we propose a spatial attentive image aesthetic assessment model to address that factor. Our method exploits the attention mechanism to learn the spatial attention map and aggregate the learned features according to it. This method preserves the image aspect ratio intrinsically, which is vital for image aesthetics as the ratio distortion usually degrades aesthetic evaluation. Numerical experimental results show that the proposed method gets the highest correlation performance with human annotation on a public benchmark, outperforming the existing state-of-arts.