Modeling Human Perception for Image Aesthetic Assessme

Deep convolution network (CNN) has shown promising progress in image aesthetic assessment. However, the properties of the network have become the bottleneck in this task. The parameter-sharing mechanism in CNN does not consider the difference between foreground and background, which is crucial for human to judge images and decide their aesthetic scores. In this paper, we proposed a Visual Perception Networks (VP-Net) to support learning from attractive region representation. We find out attractive region by introducing an unsupervised region search algorithm and conduct adaptive aggregation. Experimental results on a large scale aesthetic dataset (AVA) demonstrate the effectiveness of our approach.