High-level semantic photographic composition analysis and understanding with deep neural networks

Abstract

In order to take better photos, it is a fundamental step for the beginners of photography to learn basic photo composition rules. However, there are no tools developed to help beginners analyze the composition rules in given photos. Thus, in this study we developed a system with the capability to identify 12 common composition rules in a photo. It should be noted that some of the 12 common composition rules have not been considered by the previous studies, and this deficit gives this study its significance and appropriateness. In particular, we utilized deep neural networks (DNN) to extract high-level semantic features for facilitating the further analysis of photo composition rules. In order to train the DNN model, our research team constructed a dataset, which is collected from some famous photo websites, such as DPChallenge, Flicker, and Unsplash. All the collected photos were later labelled with 12 composition rules by a wide range of raters recruited from Amazon Mechanical Turk (AMT). Two DNN architectures (AlexNet and GoogLeNet) were then employed to build our system based on the collected dataset. The representative features of each composition rule were further visualized in our system. The results showed the feasibility of the proposed system and revealed the possibility of using this system to assist potential users to improve their photographical skills and expertise.