I primarily conduct research in embedded signal and image processing systems. An embedded system is a part of a product that works "behind the scenes". When a person speaks into a cellular phone, the cellular phone converts voice (a signal) into a radio frequency transmission. This conversion is carried out by a series of systems including a microphone, voice digitizer, and digitized voice signal compressor. Each of these systems is called an embedded system, and the digitized voice signal compressor is an example of an embedded digital signal processing system. Similarly, when a person takes a picture with a digital still camera, an optical signal is digitized, enhanced, and compressed by a series of embedded systems. Compression enables more images to be stored in the same amount of memory with little or no loss in visual quality.
Designers of embedded digital signal and image processing systems try to optimize the tradeoff between signal quality and implementation complexity, subject to constraints on the physical size, power consumption, and standards compliance. Implementation complexity includes execution time and memory used and has been studied by Computer Scientists and Engineers. Signal quality can be measured in communication systems in the form of bit error rate (how often is a bit flipped from 0 to 1 or from 1 to 0) and bit rate (how fast can data be sent). The perceived signal quality of audio, image, and video can be measured by using methods in the psychophysics literature. In optimizing tradeoffs, designers apply linear systems theory, graph theory, and optimization theory in algorithm analysis, modification, and development. (In a linear system, doubling the input value would double the output value.) Designers model algorithms in electronic design automation programs to simulate their behavior and to convert them into hardware and/or software implementations to estimate physical size and power consumption. By its nature, embedded systems design is inherently multidisciplinary.
A ubiquitous embedded implementation technology is the programmable digital signal processor. Several programmable digital signal processors are in each PC, e.g. in the sound card, disk drive, fax/modem card, and CD player. Programmable digital signal processors are available as chips that can be plugged into printed circuit boards, and as cores that can be put into an integrated circuit design. In 2004, the programmable digital signal processor chip market had $8B revenue, according to Forward Concepts. All five major vendors of programmable digital signal processor chipsets, which collectively represent about 80% of the total market, are either based in Texas (Freescale and Texas Instruments) or design their most recent processors in Texas (Agere, Analog Devices, and Intel). Two-thirds of programmable digital signal processor chips are used in wireless handsets such as cell phones. About 10% of the chips are used in DSL and cable modems for high-speed Internet access. Other communication companies having significant presence in Texas include integrated chip and reference design makers Silicon Labs, and StarCore; product manufacturers Cisco, Dell, Lucent, Motorola, National Instruments, and Nokia; and service providers AT&T, MCI, SBC, Time Warner, and Verizon.
My research spans theoretical analysis and fast algorithm development to design and real-time implementations of signal and image processing systems. Next, I divide my research into two categories: (1) embedded signal processing, and (2) embedded image processing. In signal processing, I have been focusing on the design and real-time software implementation of asymmetric digital subscriber line (ADSL) transceivers for high-speed Internet access for residences and small businesses over leased phone lines. Worldwide, in 2004, ADSL has a 2:1 advantage in subscribers over cable modems. ADSL systems carry data on multiple harmonic frequencies, which is known as multicarrier modulation. Among the digital systems in an ADSL receiver, the equalizer has the largest effect on bit rate (connection speed). The equalizer compensates for the damage to the data incurred during transmission. I have developed two equalization structures and many online equalizer design methods to provide a variety of bit rate vs. implementation cost tradeoffs. My research results have been incorporated into multicarrier wireline transceivers by at least two different companies.
In image processing, I focus on the design and real-time software implementation of high-quality halftoning for desktop printers and smart image acquisition for digital still cameras. Image halftoning converts an image with a full range of colors to an image with a limited range of colors for printing and display. Many desktop printers, for example, can either place an ink dot on the page or not. For a black-and-white printer, image halftoning would create a binary image of black and white dots to give an illusion of shades of gray. The worldwide printer market in 2004 is estimated at $15B. My group's primary contribution is in the design, analysis, and quality assessment of halftoning by error diffusion for real-time processing by printer pipelines.
By signal processing, I mean the processing of streams of data. The data could be a stream of speech or audio samples, as on an Internet phone or in an audio CD player, or a stream of transmitted bits, as in an ADSL transceiver or cell phone. In particular, I have developed theory and embedded implementations of signal processing systems for wired and wireless communication applications.
Wired Communication Systems
A significant impact in embedded signal processing systems my group has made is in ADSL transceiver design. A transceiver is a single product (often a single chip) that resides with a modem and transmits and receives information. A transceiver provides two-way communication. An example of one-way communication is FM radio. In FM radio, a radio station transmits its audio signal in a large radius, and FM receivers play the received audio. The FM receiver does not transmit back to the FM radio station. FM radio is also an example of transmitting information using a single carrier frequency, a.k.a. the radio station frequency (e.g. 90.5 MHz).
An ADSL signal carries digital information on multiple harmonic frequencies over a wire. One can think of each carrier frequency supporting a different dial-up modem connection. With this view in mind, ADSL communication could be viewed as simultaneously using 26 dial-up modems to transmit data from the user to the telephone company (e.g. SBC in Austin), and 224 dial-up modems to transmit data the telephone company to the user. The uneven data rates, which give rise to the ôasymmetricö modifier to DSL, are a natural match to data rates required for a user who is browsing the Internet to download audio, images, and video. That is, a higher data rate is coming from the Internet service provider to the user than that from the user to the Internet service provider.
ADSL transmission is over existing telephone lines using frequencies from 25 kHz to 1.1 MHz. Existing telephone lines were designed and deployed to carry voice over frequencies from 0 to 3.4 kHz. As a consequence, ADSL transmission faces a wide variety of impairments, including echoes from unterminated splices of the wire near or on the user premises, electromagnetic coupling with other weakly shielded telephone lines in the same bundle, and AM radio stations that occupy the frequency band from 550 to 1710 kHz. It is critical for the equalizer in the ADSL receiver to compensate for these and other sources of degradation experienced by the transmitted signal.
In the following paper,
G. Arslan, B. L. Evans, and S. Kiaei, "Equalization for Discrete Multitone Receivers To Maximize Bit Rate", IEEE Transactions on Signal Processing, vol. 49, no. 12, pp. 3123-3135, Dec. 2001.
my coauthors and I presented the Minimum Intersymbol Inteference (ISI) design method for a conventional ADSL equalizer. The conventional ADSL equalizer consists of a filter, Fourier series, and scaling. (The transmitter uses an inverse Fourier series to combine independent bit streams using multiple harmonic carrier frequencies, and hence, the receiver uses the forward Fourier series to extract the independent bit streams.) The filter is known as a time-domain equalizer (TEQ). The Minimum ISI method was the first method that computes the TEQ coefficients to maximize a measure of bit rate and that could be implemented completely in real time in software on a programmable digital signal processor (which is already present in many ADSL transceivers). We released implementations on Texas Instruments and Motorola (now Freescale) digital signal processors in 2001. The above paper also presents a Maximum Bit Rate Method that maximizes a more accurate measure of bit rate using offline optimization.
What made the work particularly accessible was our development and release of Matlab toolbox for discrete multitone (DMT) TEQ design. The initial toolbox in Fall 2000 contained 10 time domain equalizer design methods for comparison using many figures of merit. The current release (3.1) contains four equalizer structures and nearly 20 design methods. More than 20 companies (including Texas Instruments and Alcatel) and 20 universities (including CalTech, Cornell, and KU Leuven) have been using the toolbox.
Many groups built on our Minimum ISI and Maximum Bit Rate methods, e.g. Rick Johnson's group at Cornell and P. P. Vaidyanathan's group at CalTech. The paper's Minimum ISI method was adopted in collaborative work with Dr. Arthur Redfern at Texas Instruments on the dual-path time domain equalizer structure:
M. Ding, A. J. Redfern, and B. L. Evans, "A Dual-path TEQ Structure For DMT-ADSL Systems", Proc. IEEE Int. Conf. on Acoustics, Speech, and Signal Proc., May 13-17, 2002, vol. III, pp. 2573-2576, Orlando, FL.
I also collaborated with Dr. Lloyd Clark (TI Geomatics) on iterative refinement approaches for multicarrier wireline equalization:
G. Arslan, B. Lu, L. D. Clark, and B. L. Evans, "Iterative Refinement Methods for Time Domain Equalizer Design", EURASIP Journal on Applied Signal Processing, Special Issue on Advanced Signal Processing Techniques for Digital Subscriber Lines, special issue no. 7, 2006.
Iterative refinement adapts to the available computation resources in computing answers (time domain equalizer coefficients in this case).
Our work continues in ADSL transceiver design, as described at:
http://www.ece.utexas.edu/~bevans/projects/adsl
Wireless Communication
Systems
Since Fall 2002, I have been extending my research in the analysis and design of wireline multicarrier transceivers to wireless multicarrier transceivers. Wireless multicarrier transmission is used by several recent wireless Internet access standards, e.g. IEEE 802.11a and 802.11g for wireless local area networks and IEEE 802.16 for metropolitan access networks. IEEE 802.11a/g systems have been finding their way onto laptops. IEEE 802.16, which is for fixed antenna systems deployed in the residence and small business, has been in field trials.
In Fall 2002, I was on faculty research assignment at Cornell University to work with Prof. C. Rick Johnson and his Blind Adaptive Receiver Design research group. Together, our groups have published two journal papers and have a third one accepted for publication. Our first journal paper together has dual application to wireless and wireless multicarrier systems:
R. K. Martin, M. Ding, B. L. Evans, and C. R. Johnson, Jr., "Infinite Length Results and Design Implications for Time-Domain Equalizers", IEEE Transactions on Signal Processing, vol. 52, no. 1, pp. 297-301, Jan. 2004.
In this paper, we showed that ôlonger is not always betterö when it comes to equalizer design for a multicarrier communication system. In particular, we showed that after a certain point, the adding of free parameters (coefficients) to the equalizer filter actually causes the communication performance to degrade. The collaboration with Prof. Rick Johnson continues to the present time.
In 2003, Prof. Heath and I combined my research in time domain equalization (channel shortening) with his research in multi-antenna systems to develop joint space-time equalizers. These equalizers shorten the effective channel impulse response while maintaining the diversity advantages of multiple antennas at the transmitter and receiver. The net effect is to increase achievable data rates and decrease the complexity of the decoder that follows the equalizer.
Since Fall 2003, Prof. Andrews and I have been collaborating on a variety of design issues facing the next generation of wireless multicarrier communication systems. In current wireless multicarrier communication systems, only a single user can transmit on all of the carrier frequencies at any given time. To support multiple users, a base station assigns each user its own time slot or frequency band to use in exchanging data with the base station. The major drawback to this static allocation of resources is the fact that different users experience different wireless channel effects and this experience changes over time. For one user, a particular carrier frequency may be unusable, but for another use, the same carrier frequency may provide a high data rate transmission. The opposite might be true some time later.
One of the solutions proposed in the next generation of wireless multicarrier communication systems, which is embodied in the emerging IEEE 802.16a standard, is to allow all users to transmit data at the same time but use different subsets of carrier frequencies. These subsets are not necessarily contiguous. The goal is to increase the overall system data rate. Prof. Andrews and I have derived the mixed-integer optimization problem. We have also developed heuristics for near-optimal solutions that are linear in time and space with respect to the number of carrier frequencies and the number of users.
Our work continues in OFDM system analysis and design, as described at:
http://www.ece.utexas.edu/~bevans/projects/ofdm
With the explosion of the popularity of the World Wide Web and the growing digital still camera market, consumers are commonly acquiring, displaying, enhancing, and printing digital images. Browsing compressed images in JPEG, TIFF, and other formats is simply part of browsing the Web. The process of printing an image converts each grid point in the image (called a pixel) into one or more printed dots (halftone image). My research in image processing has focused on three key areas: áimage acquisition, image compression, and image halftoning.
Image Acquisition
My research in image acquisition applies in two different imaging modalities: optical imaging and radar imaging. The radar images of the earthÆs surface are acquired from airplanes. In collaboration with Prof. Melba Crawford, my group has developed new low-complexity algorithms to improve the vertical resolution of surface elevation measurements in the radar images. The improved resolution is vital for flood control planning in relatively flat areas.
I develop a framework for automating photographic composition rules used by professional photographers. The key step is to locate the main subject. The framework controls the optical settings to take a supplementary picture in which the objects not in the plane of focus are blurred, thereby leaving the main subject in focus. Then, a digital image processing algorithm roughly extracts the main subject based on edge enhancement, edge detection and contour smoothing. The main subject location algorithm does not require prior knowledge of the scene content or indoor/outdoor setting. The supplementary image is then registered with the image that the user took. Using the framework, my group has developed algorithms to automate three photographic composition rules: (1) placement of the main subject obeying the rule-of-thirds, (2) background blurring to simulate the main subject being in motion or decrease the depth-of-field of the picture, and (3) merger detection and mitigation when equally focused main subject and background objects merge as one object. The algorithms to locate the main subject location and automate the three photographic rules take one pass over the supplementary image to reduce the memory input/output through the processor and can be implemented in fixed-point arithmetic on the type of programmable digital signal processor found in many digital still cameras.
Our work continues in image acquisition, as described at:
http://www.ece.utexas.edu/~bevans/projects/dsc
Image and Video Compression
Image and video compression is of particular interest in transmission, e.g. on the World Wide Web. Image compression is also used in digital still cameras and desktop printers to increase the number of images stored in memory. Video compression is also used in digital cable TV, high-definition TV, and DVD. In collaboration with Prof. Alan Bovik, I have developed real-time wireless video coding and decoding systems. These codecs exploit properties of the human visual system to reduce the bit rate (file size) for the same visual quality. In collaboration with Prof. Gustavo de Veciana, my group analyzed the network traffic in packets produced by a variable bit-rate MPEG-4 encoder. We found that the network traffic was fractal, i.e. self-similar, which can aid in designing video servers for the Internet. Our work continues in image and video compression, as described at
http://www.ece.utexas.edu/~bevans/projects/dsc
Image Halftoning
The operations in printing a high-resolution image (e.g. a photograph or picture from a digital still camera) involve compression, decompression, color conversion, resizing/interpolation, and halftoning. Prior to halftoning, the image retains high resolution in both intensity and color. Halftoning reduces the resolution in intensity and color, e.g. a 24-bit color image to a three-bit color image or an 8-bit grayscale image to a binary image, for printing and display. Until the late 1990s, printing presses, ink jet printers, and laser printers were only able to apply or not apply ink to paper at a given spatial location. For grayscale printing, the ink dots were black. For color printing, a cyan, magenta, and yellow ink dot is possible at each spatial location. Many color printing devices can also produce black ink dots. Low-cost liquid crystal displays (LCDs) have the same limitation in that they can only turn a pixel on or off.
Halftoning is more complicated than simply truncating a multi-bit resolution to a smaller resolution. Simple truncation would give poor image quality because the quantization error would spread equally over all spatial frequencies. Instead, binary halftoning would try to compute a pattern of binary dots to achieve the illusion of a multi-bit image. One way to achieve the illusion is to push the quantization error into high frequencies where the human visual system is less sensitive. Examples include error diffusion halftoning (which employs feedback filtering of the quantization error) and direct binary search (which employs an iterative algorithm on the halftone to make it look more like the original image to the human visual system).
In printers, fax machines, and copiers, halftoning is the final step before a document is printed. All three systems have real-time constraints in pages per minute, so any processing must be fast and require little memory. Laser printers, for example, process 220 MB/s. Book printers process 7,344 MB/s. For reference, the data rate for broadcast video is about 24 MB/s. In the high-volume, low-cost printer market, error diffusion halftoning is used in many color inkjet printers and some laserjet printers. Direct binary search is too computationally intensive to be implemented in a printer, copier, or fax machine.
For color halftoning by error diffusion, I have developed (1) a unifying theoretical framework to describe its behavior, (2) methods to compensate the image distortion it induces, and (3) methods for subjective and objective halftone quality assessment. The theoretical framework is based on modeling the nonlinear error diffusion system as a linear system by replacing the color quantizer with a matrix gain plus additive uncorrelated noise source. Linear methods to compensate the image distortion include vector-valued prefiltering to invert the signal transfer function and vector-valued adaptive filtering to reduce the visibility of color quantization noise. The visibility of color quantization noise is based on an objective measure that utilizes a linear model of the color quantizer and a linear model of the color response in the human visual system. Another method to compensate false textures in the halftone (i.e. textures that are not visible in the original) is to replace the quantizer with a lookup table that flips the outcome near threshold values. All compensation methods are of low enough complexity to be incorporated into a printer pipeline. A subjective quality assessment framework for color halftone display on PC monitors was conducted in a controlled laboratory setting and in a Web-based framework, using the sRGB monitor standard. The quality assessment was based on a force choice test given an original image and two candidate halftones. The subjective testing was used to determine which of four linear color spaces gave the best subjective color halftone results.
Our work continues in image halftoning, as described at
http://www.ece.utexas.edu/~bevans/projects/halftoning
My research spans theoretical analysis and fast algorithm development to design and real-time implementations of signal and image processing systems. In signal processing, I have been focusing on the design and real-time software implementation of ADSL transceivers for high-speed Internet access. In transceiver design, my group's primary contribution is the first ADSL equalization structure that maximizes a measure of bit rate and is realizable in real-time fixed-point software. In image processing, I focus on the design and real-time software implementation of high-quality halftoning for desktop printers and smart image acquisition for digital still cameras. In imaging, my group's primary contribution is in the design, analysis, and quality assessment of halftoning by error diffusion for real-time processing by printer pipelines.