Photo Cropping via Deep Reinforcement Learning

Automatic image cropping aims at changing the composition of images to improve the aesthetic quality of images. It can provide professional advice for image editors and save time. Most of the existing automatic image cropping methods are based on specific features such as aesthetic features or salient features. These methods adopt sliding window mechanism to generate numerous cropping candidates, and then select the final results based on these specific features. It is very time-consuming and can only produce cropping results of a limited aspect ratio. In the face of these situations, a DLRL (deep learning framework combined with reinforcement learning) framework is proposed for image cropping, which only uses the basic features of the image for cropping without producing numerous candidate windows. Moreover, cropping step by step is more in line with the process of image cropping by people using Photoshop or other software. Experiments show that the proposed method can save a lot of time and improve cropping efficiency. The method proposed achieves the state-of-art performance in the open Flickr Cropping Dataset and CUHK Image Cropping Dataset.