摘要 |
A method and camera apparatus touches up a source image to produce a target image. The source image is partitioned into non-overlapping tiles of pixels. Each tile is labeled. A probability distribution of the labels is inferred, in which the probability distribution is a conditional random field. Weights are determined from the conditional random field. Then, each tile of the source image is transformed according to the weights to produce a corresponding tile of a target image. The transforming maximizes a conditional likelihood of the target image given the source image, while marginalizing over all possible labelings of the source image.
|