Gradient domain photomontage via interactive object selection

Authors: BEKİR DİZDAROĞLU

Abstract: One of the most critical steps in photomontage is marking an object that is to be copied from a source image to a target image efficiently. Some unexpected and unwanted effects such as bleeding usually occur on the resulting image if the object is roughly marked, copied, and then pasted to the target region, even when using the Poisson equation approach. Moreover, original color information for the pasted object cannot be preserved, and color values for the boundary pixels of the pasted object may change and/or blurriness may occur when using some methods described in the literature. The proposed methods presented in this paper are designed to efficiently overcome these problems. The problem of rough selection is solved by employing the modified intelligent scissors method. In order to prevent diffusion around the boundary of the pasted object, a band region is created, and the texture information from the source image is used. In order to better preserve the color information of the pasted object, a fidelity term is added to the Poisson equation. Our approach for producing a photomontage is more efficient than other methods in the literature in terms of interactive object selection and automated band region creation. In our approach, readjustment is not necessary for required parameters in the operations. Qualitative and quantitative results prove the advantages of the proposed methods.

Keywords: Band region, fidelity term, finite differences, gradient vector field, interactive object selection, photomontage, Poisson equation, spectral method

Full Text: PDF