MLANs: Image Aesthetic Assessment via Multi-Layer Aggregation Networks

Image aesthetic assessment aims at computationally evaluating the quality of images based on artistic perceptions. Although existing deep learning based approaches have obtained promising performance, they typically use the high-level features in the convolutional neural networks (CNNs) for aesthetic prediction. However, low-level and intermediate-level features are also highly correlated with image aesthetic. In this paper, we propose to use multi-level features from a CNN for learning effective image aesthetic assessment models. Specially, we extract features from multi-layers and then aggregate them for predicting a image aesthetic score. To evaluate its effectiveness, we build three multilayer aggregation networks (MLANs) based on different baseline networks, including MobileNet, VGG16, and Inception-v3, respectively. Experimental results show that aggregating multilayer features consistently and considerably achieved improved performance. Besides, MLANs show significant superiority over previous state-of-the-art in the aesthetic score prediction task.