Pictured: Liu’s fractal-based scheme for detecting cancer in digital mammograms. (left) The original mammogram, (center) segmented, and (right) with cancer detected.

Advances in Digital Signal Processing (DSP) are the key to the future of communications. Whether we are talking on the phone, browsing the internet, or watching TV, chances are that the signals we are receiving or sending are at some point being processed by a computer or digital processor. This processing usually means the signal is being maximized for clarity and speed—either on the sending end, the receiving end, or both. So as more and more data continues to travel through the same data lines, better ways of compressing and transmitting data must be developed, as well as faster and more efficient systems and paths to send them through.

Increasing and enhancing the efficiency and quality of digital signal processing is a major focus of research at Maryland. Prof. K. J. Ray Liu (EE/ISR), has been leading the research thrust in this emerging and important area since 1990. Since establishing the Digital Signal Processing Laboratory (DSPLAB) in 1993, Liu has directed a group of about 20 Ph.D. students and postdoctoral researchers working in three important areas: image and video technology, signal processing for wireless communications, and the design and implementation of signal processing systems.

Image and Video Technology

Image and video data take up lots of bandwidth, so researchers are challenged with the task of trying to compress these types of data and improve the systems they are transmitted through so they are sent and received clearly, quickly, and with little power consumption. In this research area, Liu has developed: 1) low-complexity, real-time systems for video coding; 2) combined source coding/channel coding/modulation for efficient image/video transmission; and 3) advanced medical imaging technology.

Video data is commonly pared-down and compressed using video coding/decoding schemes, or Codecs. These Codecs typically utilize a discrete cosine transform (DCT) algorithm for removing spatial redundancy (between pixels) and motion estimation (ME) for removing temporal redundancy (between successive video frames). Significantly, DCT and ME are generally considered distinct, resulting in high latency and complexity. Liu and researcher U.V. Koc (Ph.D. ‘96, now with Bell Labs) proposed combining DCT and ME into a single component—which, when coupled with Liu’s reduction of the ME complexity, results in a Codec that is both faster and simpler than existing schemes. This combined DCT/ME approach, by eliminating the need for interpolation, simplifies the implementation of sub-pixel ME. This approach, developed by Liu and researcher J. Chen (Ph.D. '98, now with Bell Labs), can be easily applied to new multimedia standards such as MPEG-4.

Multimedia communications will soon enter the wireless domain, but with current problems such as limited bandwidth, multipath signals and channel fading preventing the guarantee of high-quality multimedia communications, researchers are focused on developing new schemes for providing robust, low-complexity, low bit-rate, and low-power processing for multimedia data communications under a bandwidth-limited constraint.

Liu and researcher H. Zheng are investigating the transmission of image and video data based on unequal error protection, multi-carrier modulation, multiresolution coding and modulation. They recently developed a new loading algorithm using multicarrier modulation for layered, coded source transmission. They also considered a subband modulation approach to minimize data distortion over noisy channels by proper coding, modulation, and mapping from source codeword to channel modulation symbols.

The medical field is another important area for image processing. Liu and researcher H. Li (Ph.D. '97, now with Odyssey Technologies) have developed a fractal-based scheme that can separate an X-ray image of normal body tissue from cancerous tissue in digital mammograms. Their approach eliminates the conventional reliance on prior knowlegde of suspicious regions. His group also employs complex statistical image models and neural networks to enhance and detect even more difficult cancer patterns. This process is depicted in Fig. 1. Liu’s cancer-detection scheme works without distorting or destroying any of the disease’s features, and is ideal for computer-aided breast cancer detection.

DSPLAB researchers have also investigated a wavelet-based multiresolution approach to X-ray tomography. While the traditional implementation of tomography requires taking a large X-ray for the entire cross-section of a sampled region, Liu and researcher F. Rashid-Farrokhi (Ph.D. '97, now with Bell Labs) proposed a solution that needs only a small, localized X-ray to achieve similarly accurate results. This reduces a patient's exposure to X-rays, and because the actual sample is smaller, the image reconstruction speed is fast enough for real-time surgery applications.

Signal Processing for Wireless Communications

Resarchers are trying to find ways to combat the limited capacity of wireless networks, especially as the number of wireless users continues to grow. Liu’s research in this area focuses on developing efficient signal processing algorithms for optimizing system performance. This approach jointly involves coding, equalization, power and rate control, and diversity and beamforming.

Smart antennas are one type of technology to receive recent attention in wireless research. These devices utilize DSPs to manage all of the switching and management of frequencies, increasing capacity sometimes by as much as 30%. They also offer the advantage of space diversity.

In a wireless network, users need a basestation that provides a quality signal using a minimum amount of power. Liu and Rasid-Farrokhi discovered that if smart antennas “cooperate” with one another, power can be reduced even further, so they developed a combined smart antenna, power control, and basestation assignment approach that reduces the required transmitted power by one order of magnitude, enabling a larger user capacity. In addition, Liu and J. Razavilar (Ph.D. '98, now with 3Com) studied the throughput rate and latency affected using smart antennas. Further research, conducted with graduate student K. Vissa, combined power control with blind equalization in smart antenna systems so that no prior knowledge of antenna gains was necessary.

The smart antenna system developed by Liu’s research group will automatically adjust itself to the environment. Liu and graduate student M. Olfat explored power control schemes using smart antennas to improve the performance of Orthogonal Frequency Division Multiplex (OFDM) systems. Liu and Razavilar developed optimal rate control schemes for the selection of data rates that can maximize wireless signal throughput and minimize switching. Liu and H. Wang (Ph.D. '96, now with Bell Labs) also considered configurations of smart antennae under multipath fading.

Signals travelling through wireless channels pick up distortion. Equalization schemes are used to compensate for this distortion. Traditionally, training signals are used over wireless channels to “train” an equalizer or prepare it for the incoming signal. Under limited bandwidth conditions, however, there is no luxury for training signals. Performing equalization without training signals is called blind equalization. Liu, B. Sampath (Ph.D. '97) and Y. Li (Postdoc, '96, now with AT&T Research Labs) proposed many new blind equalization schemes, ranging from single channel to multi-channel. These algorithms have a wide range of trade-offs in terms of complexity and convergent rates. One subspace blind algorithm developed by DSPLAB researchers can converge by using only 20 to 30 symbols. Another algorithm they have refined requires very little computing power.

Broadband infrastructures such as ADSL may soon become a popular scheme for voice and data communications for many home users. The most common coding scheme used in ADSL is Discrete Multitone (DMT) modulation, a multicarrier modulation scheme that utilizes the cyclic prefix to separate subchannels from each other to reduce interference. Dr. Liu and graduate student X. Wang recently proposed to further exploit the cyclic prefix for adaptive channel estimation. As such, if the channel response is changing with time, their scheme will be able to automatically track the channel and provide the best performance.

Design and Implementation of Signal Processing Systems

An important focus in signal processing research is to improve the systems and architectures that transmit data. Liu’s research group has focused primarily on developing high-performance, low-power algorithms and architectures for adaptive filtering, image and video processing, and communications. Liu and A. Wu (Ph.D. '95, now faculty of National Central Univ, Taiwan) developed square-root and division-free recursive least squares architectures for radar and wireless applications. In addition, Liu and C.T. Chiu (Ph.D. '92, now with Bell Labs) developed high-speed transform coding architecures for HDTV. Liu and graduate students N. Chandrachoodan and O. Dikmen are investigating wavelet-based video coding systems.
Low-power design has figured prominently in wireless communications applications. There are many ways to achieve low-power design. Liu and Wu have developed a systematic, low-power and high-performance, dual-purpose, algorithm-based architectural approach for computationally intensive signal processing systems. This approach is independent of fabrication and design technology. When used in conjunction with other low-power technologies, significant power savings can be achieved. Liu, Chen, Wu and A. Raghupathy (Ph.D. '98, now with Qualcomm) have applied this design methodology to filtering, adaptive equalization, image and video coding, channel coding, and more.
For more information about Liu’s research in digital signal processing, visit the DSPLAB’s web site.

Features | Department News | Staff News | Alumni News | Faculty News | Student News