Automatic photographic composition based on Convolutional Neural Network

Abstract

This paper presents an automatic photographic composition method based on convolutional neural network. Firstly, photographic composition rules are adopted for photographic composition assessment, including the rule of third, horizontal line, leading line, scale and texture information. Then, the scores obtained by the aforementioned rules are used to train the Convolutional Neural Network (CNN). To further improve the photographic composition CNN model, a dual-channel CNN model is proposed. The top channel contains the network of scores of composition rules, whereas the bottom channel is trained by the image scores rated by the crowd. The experimental results show that the CNN model is very effective for automatic photographic composition suggestion.