Aesthetic guided deep regression network for image cropping

Cropping an image to improve its aesthetic quality is a challenging problem because aesthetic, which defines the harmony and beauty in the image, is “really in the eye of the beholder”. Even for the same image, different viewers might have various opinions of optimal composition with respect to aesthetic. To accomplish this subjective task, a deep learning framework is designed where the visual fixations of the image is detected based on selected deep representations and an initial visual saliency rectangle is generated to include the interested objects consequently. Afterwards, a cropping rectangle is proposed by mapping the initial visual saliency bounding box to optimal cropping areas through a regression network, where the relationship between interested objects and the optimal composition of the image is discovered. The experimental results on public datasets show that the proposed method has the competitive results than state-of-the-art approaches. •Learn the relationship between interested objects and optimal cropping window.•Accomplish the image cropping with high efficiency.•Investigate different saliency detection approaches for improving image cropping performance.