An Experience-Based Direct Generation Approach to Automatic Image Cropping

Automatic Image Cropping is a challenging task with many practical downstream applications. The often divided into sub-problems - generating cropping candidates, finding the visually important regions, and determining aesthetics to select most appealing candidate. Prior approaches model one or more of these separately, combine them sequentially. We propose novel convolutional neural network (CNN) based method crop images directly, without explicitly modeling image aesthetics, evaluating multiple detecting salient regions. Our trained on large dataset cropped by experienced editors can simultaneously predict bounding boxes for fixed aspect ratios. consider ratio be critical factor that influences aesthetics. automatic cropping, did not enforce outputs, likely due lack datasets this task. We, therefore, benchmark our public two related tasks first, aesthetic regard ratio, second, thumbnail generation requires but where are crucial. show strategy competitive performs better than existing methods in both tasks. Furthermore, one-stage easier train significantly faster two-stage end-to-end inference. present qualitative evaluation study, find able generalize diverse from unseen retains compositional properties original after cropping. results demonstrate visual attention regions necessarily required build algorithm.