Image composition with blurring effect from depth of field

This paper describes a new framework for image composition according to the blurring effect bred by changing depth of field from the target images. The framework involves two stages: a learning phase, in which the target image, with one part of the image purported to be "further" and another part "near", is presented as "learning data" and the learned filter is applied to some objects in source image; and a composition phase, in which those blurred objects of source image are composed into the "further" part of the target image. The framework is based on a simple multiscale Gaussian filter, inspired primarily by recent results in texture synthesis and image editing.