A multi-scene deep learning model for image aesthetic evaluation

Aesthetic evaluation of images has attracted a lot of research interests recently. Previous work focused on extracting handcrafted image features or generic image descriptors to build statistical model for aesthetic evaluation. However, the effectiveness of these approaches is limited by researchers' understanding on the aesthetic rules. In this paper, we present a multi-scene deep learning model (MSDLM) to enable automatic aesthetic feature learning. This deep learning model achieves better results because it improves performance on some major problems, including limited data amount and categories, scenes dependent evaluation, unbalanced dataset, noise data etc. Major innovations are as follows. (1) We design a scene convolutional layer consist of multi-group descriptors in the network elaborately so that the model has a comprehensive learning capacity for image aesthetic. (2) We design a pre-training procedure to initialize our model. Through pre-training the multi-group descriptors discriminatively, our model can extract specific aesthetic features for various scenes, and reduce the impact of noise data when building the model. Experimental results show that our approach significantly outperforms existing methods on two benchmark datasets.