As combat commanders plan increasingly complex missions, the need for air vehicles to acquire data, process information, and communicate escalates accordingly. Unless engineers can further compress enormous amounts of data without sacrificing information quality, the Department of Defense faces extensive modifications to existing weapons systems - both to meet demands for increased bandwidth and to exploit the broader range of data-related capabilities such expansion permits.

In response, laboratory-sponsored researchers, in conjunction with the AFRL Visiting Faculty Research Program, are producing some very exciting image processing enhancements. These improvements, a result of recent advances in applying genetic algorithms (GA) to wavelet-based images, promise not only to increase the quality of transmitted images but also to decrease associated file size as compared to the results of current file compression methods.

To address the challenge of improved image file compression, the AFRL team first defined three possible approaches: (1) decrease one- and two-dimensional (1-D and 2-D) mean squared error (MSE) while maintaining file size, (2) maintain MSE while increasing 1-D and 2-D signal compression (decreasing file size), or (3) decrease both MSE and file size. Specific efforts to explore the first of these potential approaches - decreasing signal MSE - progressed significantly in 2004 when Dr. Eric Balster and Mr. Pat Marshall (of AFRL) and Dr. Frank Moore (assistant professor of computer science at the University of Alaska Anchorage) documented their work results in a report entitled 'Adaptive Filtering in the Wavelet Transform Domain Via Genetic Algorithms.' (To view this report, access the Technical Support Package [TSP] at http://www.afrlhorizons.com, reference IF-05-11). Essentially, the researchers utilized AFRL's Embedded Information Systems Laboratory resources to develop and employ a GA to optimize inversewavelet- transformed images subjected to quantization errors.

Recently, a group of Dr. Moore's students working follow-on efforts in support of the original project achieved significant breakthroughs yielding additional improvements. One of their objectives was to determine whether they could reduce image processing time by initially training the GA on a subimage and then using the evolved transform to optimize a larger image containing either the subimage or characteristics similar to those of the subimage.

While the researchers hypothesized increased GA performance due to less computational time required to train on the smaller subimage, the results were even better than anticipated. Whereas the original GA required more than 46 hrs to process 5 images over 500 generations using a population size of 500, the students' method took approximately 70 min to find a solution based on the same number of generations and an identical population size. In just over 4 hrs, their modified GA processed 5 subimages for 200 generations using a population size of 3,000.

Pursuing a second improvement objective, the student group proposed to optimize both reverse and forward wavelet transform coefficients using the modified GA method. Once again, the modified GA performed better and faster than standard wavelets, locating a solution (in just 10 generations with a population of 50) superior to that of the original GA (which required 500 generations with a population of 200). These results conclusively prove the usefulness of evolving both reverse and forward wavelet transform coefficients.

The students' third research objective was to determine whether they could further compress image file size by adjusting standard wavelet transform coefficients using a GA. Their results revealed that researchers can indeed realize an approximate 19% compression improvement - without adversely affecting MSE results - using the evolved coefficients. The finding marks an especially significant discovery, given that researchers were previously unaware that file compression could be improved using GAs.

This collaborative program between AFRL and University of Alaska Anchorage researchers provides considerable evidence suggesting that the search space of nontraditional transforms is rich with beneficial solutions. The original bitmap (.bmp) image shown in Figure 1 exemplifies the type of dramatic results achieved with 2-D images. Researchers compressed this original image using a standard Daubechies-4 wavelet transform subjected to a quantization step of 64 and then reconstructed it using the Daubechies-4 wavelet inverse transform. Figure 2 reflects the resulting image; the grayscale intensity indicates its total MSE - as relative to the original image - is 115.80.

Next, researchers employed a GA running for 1,700 generations with a population size of 200 to evolve the coefficients describing a matched forward and inverse transform pair. They then used the evolved transform to compress and subsequently reconstruct fruits.bmp under conditions again subject to a quantization step of 64, ultimately producing the image depicted in Figure 3. Compared to the original fruits.bmp, the total MSE of the image in Figure 3 is 12.37 - a figure that represents an error reduction of 89% in relation to standard Daubechies-4 wavelet performance. The difference in image quality is quite apparent even in visual comparisons of the two pictures.

This technology can provide significant benefits to the warfighter. Aside from advantages associated with increased bandwidth, the higher-quality images equate to targets that are easier for pilots, analysts, and commanders to recognize. Moreover, analysts may be able to use the image processing system to further refine images for highlighting specific targets of concern by adjusting wavelet filter coefficients using a GA. Specifically, researchers foresee this technology helping at the operational level in at least two ways:

��� Target analysts can more easily search for specific targets of concern, such as a hooded person holding an AK-47 rifle in a crowd of people. Using GAs, analysts can optimize image processing to show the AK-47 as a very sharp and crisp subimage within the overall, blurred image. Once the analyst locates the AK- 47, he or she then has the option to revert to the original, nonoptimized image to view the entire scene.

��� Automatic target recognizers (ATR) will operate more efficiently. As analysts adjust various wavelets to emphasize certain targets, the ATR's detection and recognition algorithms will operate with much higher accuracy and trigger fewer false alarms.

Currently, the research effort to improve file compression methods is limited to processing still images. However, when the technology matures, AFRL will perform the software conversions necessary for execution on digital processing hardware for streaming video. The overall operational goal of this effort is to establish the use of GAs as the predominant methodology for generating or modifying various filter coefficient sets that automatically compensate for such factors as quantization error and data file size, consistently producing high-quality reconstruction of an original signal. The research accomplished to date represents a major step towards satisfying that goal.

In addition to those researchers already named, Lieutenant Donald Tinsley, of McConnell Air Force Base, Kansas, and Messrs. Brendan Babb and Steven Becke, students from the University of Alaska Anchorage, were instrumental to the success of this research effort. In the future, Dr. Gary Lamont and other key researchers from the Air Force Institute of Technology will assist in verifying the program's most recent results and will mathematically analyze the theoretical foundation behind this new research area.

Related Links
SpaceDaily
Search SpaceDaily
Subscribe To SpaceDaily Express

Memory Foam Mattress Review

SPACE MEDIA NETWORK PROMOTIONS Solar Energy Solutions Tempur-Pedic Mattress Comparison ... Newsletters :: SpaceDaily :: SpaceWar :: TerraDaily :: Energy Daily XML Feeds :: Space News :: Earth News :: War News :: Solar Energy News