19-01-2013, 11:54 AM
Image Segmentation
1Image Segmentation.pdf (Size: 574.05 KB / Downloads: 543)
Introduction.
The goal of image segmentation is to cluster pixels into
salient image regions, i.e., regions corresponding to individual surfaces,
objects, or natural parts of objects.
Asegmentation could be used for object recognition, occlusion boundary
estimation within motion or stereo systems, image compression,
image editing, or image database look-up.
We consider bottom-up image segmentation. That is, we ignore (topdown)
contributions from object recognition in the segmentation process.
For input we primarily consider image brightness here, although similar
techniques can be used with colour, motion, and/or stereo disparity
information.
Observations on Example Segmentations
The previous segmentations were done by the local variation (LV) algorithm
[7], spectral min-cut (SMC) [6], human (H) [11, 9], edgeaugmented
mean-shift (ED) [4, 3], and normalized cut (NC) [13, 5].
• The quality of the segmentation depends on the image. Smoothly
shaded surfaces with clear gray-level steps between different surfaces
are ideal for the above algorithms.
• Humans probably use object recognition in conjunction with segmentation,
although the machine algorithms exhibited above do
not.
• For relatively simple images it is plausible that machine segmentations,
such as those shown on p.2, are useful for several visual
tasks, including object recognition.
• For more complex images (pp. 5, 6), the machine segmentations
provide a less reliable indicator for surface boundaries, and their
utility for subsequent processing becomes questionable.
• While many segmentation algorithms work well with simple examples,
they will all break down given examples with enough clutter
and camouflage. The assessment of segmentation algorithms
therefore needs to be done on standardized datasets.
Current Goals
• Provide a brief introduction to the current image segmentation literature,
including:
– Feature space clustering approaches.
– Graph-based approaches.
• Discuss the inherent assumptions different approaches make about
what constitutes a good segment.
• Emphasize general mathematical tools that are promising.
• Discuss metrics for evaluating the results.