Grant Number : NAG5-3994
Period Covered: March 1, 1999 - May
31, 1999
Report Number : 1999-Q1
I. Technical Report
During the past quarter we have concentrated our efforts on the following issues:
1. The development of an image browser
for thumbnail size images which
supports progressive
transmission of images.
2. The development of an on-board
image compression algorithm for ACS.
3. The development of a similarity
estimation method for astronomical
images by processing
compressed data.
4. Communication with astronomers
at UNLV and elsewhere.
1. To implement a web browser that supports progressive transmission of astronomical images with different formats, various classes have been designed in C++. An image class and its child bitmap class are written.
Image classes allow us to build classes for different typesof image formats.
2. At the same time we have been working on our newly proposed method for the ACS on-board data compression. We have analyzed different encoding schemes and have defined the optimum partition of image line which yields the maximum possible compression ratio. We have also tried to minimize the overhead from the service data. Only 0.03 bits per pixel are transmitted in order to let decoder know about data packing format in the partition of line, and a maximum of 2 bits are required to inform ground server about the actual code length.
On the average, 1.03 bits/pixel of overhead is required for most cases. Using different type of images, we have achieved a compression rate of minimum 5.31 and maximum 7.94 bits/pixel. For the same set of images, Pixel-pair algorithm [1] yields a compression factor in the range of 5.87 to 10.47 bits/pixel. Our method outperforms aforementioned technique mostly on noisy images in about 1.5 bits per pixel. For smoother images, the difference is not significant,i.e., about 0.5 bits/pixel.
Taking into consideration that compression time is more critical than decompression, we provide in our method a fast packing technique that delivers data in reverse order. One additional operation is included in the decompression algorithm. The computational cost have been estimated on SG platform. Experiments have shown that our method is inferior to pixel-pair algorithm but outperforms the Rice algorithm [2].
We are currently working on the time measurements for the Intel 386 which is the operating hardware platform for ACS on-board compression. Having read [3], which is the product of a joint effort with F.R. Boffi, we have contacted Mr. M. Stiavelli and discussed with him the approach that they have taken to test the execution time of the pixel-pair method. Upon our discussion, we have decided to use the same method to test the performance of our algorithm. Consequently, we are going to be using a PC with a 486/25 processor to carry out our experiments. We are confident that the similarity in instruction set between ACS' 386/16 and our 486/25 will yield very close measurements once the time obtained by 486/25 is adjusted for comparison with the results from ACS' hardware [3].
Final results of our work will be reported at the International Conference
on Imaging Science, Systems and Technology (CISST'99), Las Vegas, June
28-July 1.
3. We have analyzed an approach for image similarity estimation for content-based image access in astronomical image databases.
We have developed a method that facilitates fast search of similar images in database by processing compressed data. The automatic indexing is another key aspect of the proposed method which enhances the conventional method of manual text description and keywords tagging. Global features such as edge points location and edge density which characterize the image structure and texture are calculated from the encoded coefficients of the wavelet transform.
The novelty of our method is that it does not depend on the type of wavelet transform applied for image compression, and it is able to process both noisy and noiseless images. We preprocess data and obtain their normalized representation so that they can be fairly compared.
We have chosen compressed data analysis and processing based on the wavelet transform because this compression scheme has shown its suitability for efficient compression andd progressive transmission. I the World Wide Web, there can be found examples of astronomical image browsers which provide users with remote viewing option of large images. The wavelet based multiresolution data storage format is used for this purpose.
If a given lossy compression technique preserves important information about data such as edges and texture, then similarity can be estimated using the preserved information in the encoded data.
An analysis of wavelet-based compression methods has shown that most of the quantization and encoding techniques developed for scientific images use the aforementioned approach.
The analysis of the astronomical images has also shown that not all images, particularly those that are captured from different instruments, can be characterized as having some texture. Therefore, we introduced the image index, which indicates the type of image structure, and developed an appropriate subroutine that performs measurements. The search and measurements are limited to the images of the same index only.
Our method executes measurement routines on the base of user defined level of proximity. User specifies percentage of similarity. User interface has to provide the appropriate tools for recording data and transferring it to the server. Different levels for each feature can be specified, or one of the features can be suppressed altogether.
Because the distance vector is defined in the space of heterogeneous features, we have developed a new approach for distance vector calculation. The distance is measured by a special metric for each feature. The image is divided into blocks of predefined size, and these blocks are compared. This comparison, as a result, yields 0's and 1's in the output. The resulting distance vector in two-dimensional feature space is, thus, represented by scalar components of zero and one. The number of components depends on the number of blocks, and the value indicates whether the blocks of compared images are similar, i.e. , both features indicate similarity. If the output for the comparison is (0,0) then a 0 is assigned to the distance vector component.
Finally, Hamming weight metric is used for similarity estimation. In our approach, a weight of zero indicates coincidence of images, and the ratio of weight to the number of blocks indicates the level of dissimilarity, which is compared with a predefined level. Additionally, calculations in our approach are reduced to processing data of particular scale of the wavelet multiresolution decomposition; different levels for edge detection are supported, i.e. strong and weak edges can be found separately, and the search is interrupted as soon as the level of dissimilarity is reached. A paper on this issue has already been drafted [4].
We also plan to study the possibility of the similarity estimation based on domain-dependent context. For this purpose, different systems for storing, analyzing, and processing astronomical images will be analyzed.
4. We have also contacted two astronomers to hear their concerns and suggestions regarding the astronomical software available today in the science community. Dr. Charles Nelson from the astronomy department of University of Las Vegas, Nevada and Dr. Jason Pinkney of the astronomy department of University of Michigan gave their input on current methods and software used in space sciences, as well as their capabilities, shortcomings and potential improvements of these packages that can help astronomers in image processing.
One of the topics that were brought up was the noise contained in astronomical images. Once the raw image is downloaded from the HST, it goes through certain cleaning procedures. However, the same image is available in two different forms to whoever requests it: a raw form and a processed form. Dr. Nelson stated that each astronomer, most of the time, prefers doing his/her own cleaning on the downloaded image. Even though sometimes they download cleaned images, this does not mean that a thorough job has been done on the image (or as meticulous as they would like) and they proceed with their own, detailed cleaning, especially on the areas of the image that they want to focus the most.
Another concern that was brought up to our attention was the following: when one pulls up the preview image in HST Archive Search, one only receives a single gray scale transfer function (or "stretch"),and one magnification. So faint/small objects that are on the image cannot always be seen. It was suggested that it would be nice to have the browser adjust this stretch interactively (like SAO Image does under the option COLOR).
Finally, another suggestion was to incorporate a WFPC2-specific tool into the browser that simplifies the visualization of the 4 CCD chips, either one at a time (by clicking a 1 2 3 or 4 button, for instance) or all at once (like the preview).
Taking into consideration the discussed issues, we are going to include the following functions into the applications being developed.
Two options for the image browser will facilitate the selection between suppressed noise and noisy image transmission. The compressed data can be used to filter out, among the stored transform coefficients, the insignificant ones. Soft and hard thresholding options will provide the user with the selection of a filtering type.
We need to store additional information along with archived data. Standard deviation for each of high-pass subbands of the multiresolution decomposition is required to support online thresholding. Only the coefficients whose values are above the threshold value will be transmitted.
Another issue is to preserve and transmit the transformed data without any possible distortions. For this purpose, we are going to implement integer to integer Haar transform, which not only prevents truncation of non-integer values but also does not need quantization, hence it is faster.
Since it is desirable for the astronomers to observe each of four WFC images separately, we are going to include the mentioned option into the image browser tools.
Since the brightness of the image object and its shape are of interest, we are going to study the techniques for automatic measurements of these features by analyzing the compressed data.
II. STAFFING ACTIVITY
No change.
III. HARDWARE, SOFTWARE, AND WWW
No change.
IV. REFERENCES
1. White R. L., Becker, I. "On-board Image Compression for the HST
Advanced Camera for Surveys", Proceedings of the SPIE conference,
March 1998, Kona, Hawaii.
2. Rice R.F., Yeh P.-S. and Miller W.H., "Algorithms for High Speed
Universal Noiseless Coding", Proceedings of the AIAA Computing in
Aerospace 9 Conference, 1993.
3. Boffi F.R., Stiavelli M., "Performance of the Onboard Compression
Algorithm for ACS", Instrument Science Report ACS-98-04, Jan. 28,
1999.
4. E. Regentova, S. Latifi,
"Image Similarity Estimation by Processing Compressed Data", draft paper.