Abstract:
In this thesis, we present a new image completion method that automates the
filling in of holes left by the removal of undesired areas in images so that the
final output image is visually plausible. The reconstruction of the hole is based
on the assumption that regions, particularly in natural images, tend to be spatially
continuous and are only separated by the hole and must therefore be linked.
Therefore, our approach is based on first creating image structure (regions boundaries)
in the hole and then propagating texture from surrounding areas constrained
by this structure. Structure reconstruction is performed in order to preserve the
global structure of the image, by creating regions in the hole with well defined
boundaries such that they match the surroundings.
The images are first segmented into homogeneous regions. The regions touching
the hole are then relabelled based on their colour and spatial distances. Similar
regions are then linked resulting in creating a new area in the hole that will
be flood-filled and then synthesised to match the surrounding structure. This
reconstructed image is then used for texture synthesis as a constraint.
Our texture synthesis method proposes two modifications to the generic texture
synthesis method and this includes a parallel synthesis order and an iterative
synthesis scheme. The parallel synthesis, in which a pixel being synthesised is
independent of other pixels during any given iteration and not affected by other
previously synthesised pixels, helps reducing the directional bias caused by sequential
scanning orders such as the raster scan. The iterative synthesis scheme allows
global randomness which will progressively converge towards fine detailed texture.
This scheme ensures that the created texture has sufficient, but not excessive, randomness
and does not have replications of entire patches. As a result, the method
is able to convert gradually the input image into plausibly synthesised image and
to remove visible boundary artifacts. The combination of the image structure and
texture synthesis methods results in having an image completion method that is
capable of dealing with images with large holes that are surrounded by different
types of structure and texture areas.
