- IP1
- RESEARCH TASK: Adaptive Encoding of Wavelet
Coefficients
PLANNED ACTIVITY
The team will be working on developing new algorithms for
sub-region retrieval which integrates integer wavelet
transforms with segmentation algorithms and entropy encoding
for better compression. In general, sub-region retrieval and
entropy encoding do not work well together because of
conflicting requirements. Variable length coding, such as
Huffman coding, preclude random access to specific locations
in the compressed file. One possibility is to split the image
into non overlapping blocks, each of which are then
appropriately encoded. The other alternative is to use image
segmentation to partition the image and then encode each of
the regions independently. This would facilitate ``object'' based
retrieval as in the case of content-based browsing. The team
will be studying these issues further.
ACTUAL ACTIVITY
Adaptive coding was not investigated further as our initial results
indicated a more ``development'' type of work and because of some
overlap with the other groups within ADL. Instead, the focus was
diverted to a new framework for compression based on Average
Interpolation Subdivision (AIS).
- IP2
- RESEARCH TASK: Error Handling Techniques
PLANNED ACTIVITY
In addition, although the initial results intented to assess the
performance of the error concealment idea appear promising,
there are various issues which need to be investigated further.
One drawback of the initial work is that it can not handle
error rates that are larger than one percent.
To improve performance for this case, a two-step implementation
may be preferred where the LL-band is investigated separately.
In other words, we could make use of the fact that this band
resembles the original image and apply a similar model-based
detection/concealment technique to remove transmission errors
before combining it with corresponding detail signals. Such a
step will reduce the number of corrupted LL-band coefficients
and, hence, alleviate the problem of hard-to-detect errors when
corrupted coefficients from different bands spatially coincide.
Future work will also concentrate on packet loss, a problem
encountered when using wireless transmission or ATM networks.
One of the main goals is to eventually combine wavelet transform
and image segmentation such that an image transmission method
can be developed which is information-based rather than
data-oriented. It is hoped that such an approach to data transmission
might make error concealment easier and even more effective.
ACTUAL ACTIVITY
A paper on this topic was presented at the IEEE International
conference on Image Processing, Santa Barbara, 1997.
N. Strobel, S. K. Mitra and B. S. Manjunath, ``Model based detection
and correction of corrupted wavelet coefficients,'' Proc. 1997
ICIP, vol. 1., pp. 925-928, Santa Barbara, October 1997.
- IP3
- RESEARCH TASK: Development
of a Texture Thesaurus
PLANNED ACTIVITY
The recent progress made by the team in developing a new image
segmentation algorithm opens new research directions in content
based retrieval. The team will investigate the application to
multispectral images and color airphotos. The team is also
collaborating with the U. Illinois DLI to evaluate the texture
features and segmentation algorithms on a much larger image
database, perhaps a thousand airphotos and other satellite
imagery. Currently, many of the airphotos are being scanned
at the MIL. In addition, ground truth for the data is being
collected at U. Arizona (Prof. Hsin-Chun Chen), where a user
interface is under development for manually annotating the data.
When completed, the visual thesaurus will be able to combine
both semantic and texture information for searching the database.
ACTUAL ACTIVITY
A demo of a thesaurus with annotations is being developed at Arizona by
Prof. Chen's group. The Netra demo was shown ath the NSF technology
exhibits center during Spring/Summer of 1997.
The texture thesaurus paper appeared in a recent issue of JASIS:
W-Y. Ma and B. S. Manjunath, ``A Texture thesaurus for browsing large aerial
photographs,'' Journal of the American Society for Information Science
(Special issue on Artificial Intelligence Techniques for Emerging
Information Systems Applications), pp. 633-648, vol. 49, No. 7, 1998.
- IP4
- RESEARCH TASK: Spatial Relationships
PLANNED ACTIVITY
An immediate consequence of automated segmentation is that it would
facilitate development of algorithms where spatial relationships between
various segmented regions can be specified. For example, it is feasible
to form queries such as ``retrieve all images with mountains and ocean''.
The visual thesaurus provides a conceptual framework to integrate
image features and image annotations.
ACTUAL ACTIVITY
This is on-going work. We are working closely with
scientists from JPL in developing a demonstration of a ``texture pair
dictionary''. We hope to have some results by the end of summer 1998.
- IP5
- RESEARCH TASK: Learning Similarity
PLANNED ACTIVITY
Another interesting research direction is to develop new algorithms
for texture feature dimensionality reduction while preserving the
perceptual similarity. The objective is to compute a (perhaps
non-linear) projection transformation that maps to a much lower
dimension space (for example, from the 48 dimensional feature
vector to a 3-dimensional space) with the property that
perceptually similar textures form clusters.
This would allow the search for matching images to proceed in
a space of much smaller dimension. Another advantage of such a
map would be a smooth transition between clusters, unlike the
rigid region boundaries used with the hybrid neural network
approach. The search for images matching a particular query
could be carried out in gradually expanding spheres until
all similar images are retrieved. Toward these ends, an initial
investigation using the Brodatz image texture library has
yielded very promising results that warrant further research.
ACTUAL ACTIVITY
Preliminary results on dimensionality reduction using
multidimensional scaling appears quite promising. A paper on
this topic was presented at the IEEE Intl. conf. on Image
Processing. An MS thesis is under preparation. Ref:
M. Beatty and B. S. Manjunath, ``Dimensionality reduction using
multidimensional scaling for image search'', Proc. IEEE
International Conference on Image Processing, Santa
Barbara, California, Vol. II, pp. 835-838, October 1997.
