Learning to resize image

Abstract

Content-aware image resizing has been a promising theme in the communities of image processing and computer vision. To the best of our knowledge, most existing methods for image resizing are unsupervised. These unsupervised methods may either fail to protect the interesting regions or cause distortion of the image structure. This paper presents a novel learning based method for seam carving by incorporating the learned boundary of the important content. Specifically, a novel boundary model of the region of interest (ROI) is learned on a set of training images at first. Then the boundary of an input image is utilized as a key prior in performing seam carving to obtain the target image. The proposed method for image resizing can generate much less seams cutting through the ROI compared with previous efforts toward the same goal. Thus, the desirable regions can be preserved in the target image and the structural consistency of the input image is naturally maintained. Experiments on two publicly available data sets demonstrate the effectiveness of the proposed method.