Segmentation by adaptive prediction and region merging

This paper presents a segmentation technique based on prediction
and adaptive region merging.
While many techniques for segmentation exist, few of them are suited for
the segmentation of natural images containing regular textures defined
on non-rectangular segments. In this paper, we propose a description of
regions based on a deconvolution algorithm whose purpose is to remove
the influence of the shape on region contents. The decoupling of shape
and texture information is achieved either by adapting waveforms to
the segment shape, which is a time-consuming task that needs to be
repeated for each segment shape, or by the extrapolation of a signal to
fit a rectangular window, which is the chosen path.
The deconvolution algorithm is the key of a new segmentation technique
that uses extrapolation as a prediction of neighbouring regions. When the
prediction of a region fits the actual content of a connected region reasonably
well, both regions are merged. The segmentation process starts
with an over-segmented image. It progressively merges neighbouring regions
whose extrapolations fit according to an energy criterion. After
each merge, the algorithm updates the values of the merging criterion
for regions connected to the merged region pair. It stops when no further
gain is achieved in merging regions or when mean values of adjacent
regions are too different. Simulation results indicate that, although our
technique is tailored for natural images containing periodic signals and
flat regions, it is in fact usable for a large set of natural images.