Deep Neural Framework with Visual Attention and Global Context for Predicting Image Aesthetics

Computational inference of aesthetics has recently become a hot topic due to its usefulness in widely applications such as evaluating image quality, retouching and retrieving image. Owing the subjectivity this problem, there is no general framework predict aesthetics. In paper, we propose deep neural with visual attention module, self-generated global features hybrid loss address problem. Specifically, can be any state-of-the-art convolution classification network compatible attention. Further, feature compensates for context information during training stage, guides learn similarity between predicted aesthetic scores ground-truths through fusing soft-max-entropy Earth Mover’s Distance(EMD). With above-mentioned improvements, proposed capable effectively predicting an efficient way. our experiments, release real-world dataset that contains 1,800 2K photos labeled by several experienced photographers, then provide thorough ablation study design choices better understand superiority brought each part framework, comparisons methods on fraction metrics. The experimental results two datasets demonstrate both accuracy efficiency achieve favorably performance.