반복적 크롭핑과 수평 보정을 이용한 영상 구도 개선

Abstract

This thesis presents a novel image composition enhancement framework using a repeated cropping and horizon correction based on a deep convolutional neural network. Firstly, we propose a repeated cropping technique for photo re-composition. For a given image, a trained network iteratively predicts cropping directions which make the image to have a better composition. The system can automatically and gradually crop the photo to follow specific composition guidelines, such as the rule of thirds and salient object size. We also propose a photo horizon correction method using convolutional neural network. The trained network estimates a slanted angle by learning generic features using a huge dataset. To obtain better performance, we utilize multiple-sized pooling layers to extract multi-scale image features. By combining proposed methods, an aesthetic composition of a given photo can be enhanced automatically and effectively, as shown in experimental results.