Image Aesthetic Evaluation Method Based On Convolutional Neural Networks

In recent decades, with the rapid development of digital cameras and networks, vast amounts of images are growing on the internet. Being an important medium of information transmission, image aesthetic quality gradually gained more and more attention. Therefore, the image aesthetic quality evaluation gradually developed into a research topic, and had been applied in many other important areas, such as image retrieval, image sharing and personal photo album management.Traditional image aesthetic evaluation methods extracted discriminate features from the raw data of datasets, and used these effective features to train a classification model. These traditional methods had achieved relatively good classification performance, however, these features were specially designed for a particular database, and it was difficult to design features that can be applied to multiple databases effectively. To solve this problem, we adopt a deep learning method, which is a kind of feature learning method, to learn discriminate features automatically.In this paper, we adopt an effective deep learning method, named Deep Convolution Neural Networks(DCNN), to evaluate the image aesthetics. Based on traditional DCNN structures, we propose three kinds of DCNN structures in order to achieve good performance,.(1)Improved DCNN(I-DCNN) structure: This network is designed based on the classical DCNN network;(2)The Network-Paralleled DCNN(NP-DCNN) structure: The NP-DCNN network is able to learn features in different depths by paralleling several I-DCNN networks, and can also solve the over-fitting and under-fitting problems;(3) The Structure-Paralleled and Data-Paralleled DCNN(NP-DP-DCNN) structure: The NP-DP-DCNN network is designed to combine the traditional features and the raw data by paralleling input data and parallel several I-DCNN networks.Experimental results show that three proposed networks are able to achieve better classification performance than most of the traditional feature designed methods.