IMAGE PROCESSING TEAM

IP1

RESEARCH TASK: Color multispectral wavelets

PLANNED ACTIVITY

Currently, the extension of the discrete wavelet transform to color and (eventually) multispectral images is under investigation. The Team will study how different filters perform for a wide array of images and the tradeoff between between speed and accuracy. Another issue of concern is across-band correlation which could be further exploited to improve image compression. For the upcoming site-visit, color images will be our main concern.

ACTUAL ACTIVITY

(See under next research task, IP2)

IP2

RESEARCH TASK: Wavelet quantization

PLANNED ACTIVITY

This research problem is closely related to a practical phenomenon encountered when using almost any kind of transform coding scheme. To accurately express transform coefficients, a higher data precision is needed. As a result, more storage space would be required if coefficients were stored exactly. To alleviate that problem, we quantize wavelet coefficients, i.e. map them to some rounded representation to reduce storage requirements. In addition, we keep the residual error image for perfect reconstruction. Different quantization schemes imply different entropies for residual images. It remains for us to investigate how filter-dependent that phenomenon is and what quantization scheme will give best results considering grayscale, color and multispectral data.

ACTUAL ACTIVITY

For both task IP1 and IP2, which are closely related, we have developed an integer wavelet transformation that can be easily generalized to color images that addresses issues related to image quantization and compression. We expect to complete a paper on this topic soon. The code is now available and a JAVA implementation with client side reconstruction of the progressively higher resolution images is under progress.

IP3

RESEARCH TASK: Wavelets over noisy networks

PLANNED ACTIVITY

Here we concentrate on networking issues involving images represented by wavelet or subband coefficients. A problem encountered in many network environments is packet loss which results in missing wavelet coefficients. Our research objective is to find out if the retransmission of lost packets can be avoided by applying suitable image restoration procedures and, as a result, if network bandwidth can be saved by recomputing data instead of retransmitting it.

ACTUAL ACTIVITY

A paper has been submitted to a conference on this issue, and we continue to investigate this further.

IP4

RESEARCH AND DEVELOPMENT TASK: Texture features, similarity search and learning

PLANNED ACTIVITY

We will continue the investigation into Gabor feature representations for multispectral/color images. During the last few months, an interesting new direction that is being investigated is the use of learning algorithms for adaptively tuning the search process, and for capturing the notion of perceptual distance (as opposed to using the Euclidean distance for similarity matching).

ACTUAL ACTIVITY

Much progress has been made, particularly with respect to texture features and color features; A journal paper appeared in the special issue on DL in IEEE T-PAMI August 96 detailing much of the work. Code has been distributed to collaborators at the UIUC DLI project. Integration with test bed is under progress, as well as a JAVA version. Also, NSF has selected this as one of the demonstrations for the technology exhibits center.

IP5

RESEARCH TASK: Spatial proximity search

PLANNED ACTIVITY

We will further continue the texture based search by constructing ``texture objects'' (e.g. coast, ``parking lots'' etc.) and encoding the spatial relationships and location information. This would make possible answering queries like ``find all vegetation of this kind close to a coastal area''.

ACTUAL ACTIVITY

In order to perform spatial search, a critical component that is required is a robust segmentation tool. Much progress has been made in developing new segmentation algorithms. A paper is under preparation. A demonstration program has been written to work with a diverse collection of color images. Initial results are very promising and this demo is also been chosen for the NSF technology exhibits center.

IP6

RESEARCH TASK: Multi-image, minimum-redundancy transforms

PLANNED ACTIVITY

Here we will explore an integrated representation of the data. Since the data are obtained by observations of the same physical phenomenon (surface of the earth for these satellite images), except for the seasonal changes, one would expect a high degree of correlation from the same sensory data. Further, in case of Landsat images, a model based approach will be used to reduce the interband correlations. A long term objective is to develop a compact representation which will allow for the reconstruction of specific images (date/time/sensor) of a given geographic region at retrieval time.

ACTUAL ACTIVITY

Emphasis has been on developing fast and numerically stable SVD update algorithms. A journal paper has been accepted for publication.

IP7

RESEARCH TASK: Characterization of wavelet lossiness; Processing in wavelet domain

PLANNED ACTIVITY

This has two objectives: how much processing can be done on the coarse resolution data without compromising on the quality of the analysis; secondly, what processing is possible directly in the wavelet domain without reconstructing the images.

ACTUAL ACTIVITY

This task was not considered to be of sufficiently high priority to undertake during the past year.