Feedback Retargeting

Abstract

Feedback retargeting combines the benefits of two previous retargeting methods: Bidirectional similarity [1] and Shift-Map [2]. The first method may have blurry areas due to patch averaging and the latter can remove entire objects. Feedback retargeting has the sharpness of shift-map and the completeness of bidirectional similarity, avoiding the removal of salient objects. In Shift-Map retargeting the output image is made from segments of the input image, and this minimizes the forward direction of bidirectional similarity. An iterative feedback procedure is developed to take care of the backward direction, assuring that the input image can be reconstructed from the output image. This is done by using Shift-Map backwards, reconstructing the input image back from the output image. Areas in the input image that are difficult to reconstruct from the output image get a feedback priority score. A second Shift-Map retargeting is then performed, adding this feedback priority to the data term. These regions now have a higher priority to be included in the output. After a few iterations of forward retargeting and backward feedback the retargeted image includes all salient features from the input image. Computational efficiency and image sharpness remain as high as in ordinary Shift-Map.