http:^^www.eecs.umich.edu^systems^comresearch.html
来自「This data set contains WWW-pages collect」· HTML 代码 · 共 444 行 · 第 1/2 页
HTML
444 行
Date: Tue, 26 Nov 1996 00:54:35 GMT
Server: Apache/1.1-dev
Content-type: text/html
Set-Cookie: Apache=gs35624184896967582; path=/
<html><BODY bgcolor="#FFFFFF"TEXT="#000000"LINK="#0000FF"VLINK="#CC0000"ALINK="FF9933"><H1>CURRENT COMMUNICATIONS RESEARCH LABORATORIES AND PROJECTS</H1><hr><hr><br><DT><B>Information Theory and Information Retrieval<p>Graduate Students: N. Leung and S. Lawson<br>Professors: J. T. Coffey and S. Sechrest<br></B><DD>This project examines the fundamental limits of efficientcontext-based retrieval in large scale databases, such asthose used in scientific and medical databases. The natureand quantity of the data involved in these applicationsdemand new approaches, and it is the goal of this project toinvestigate the role that the methods of information theorycan play in this task. Based on a number of simplifiedabstract models, we have derived results that demonstratethat expected access time can be greatly reduced in generalby adding redundancy to the database. The general probleminvolves a number of interesting variants of classicalproblems in source and channel coding and multi-userinformation theory.<p>In developing theoretical results in this research, we areaiming to acquire insight that will be used to providefirst-order guidance in system design. Further guidance willbe found by comparisons with results arrived at through moredetailed simulation of systems. The applicability androbustness of our results and others are being investigated.Actual datasets and realistic workloads can be used tovalidate our models and assess the applicability of ourresults to physical systems.<hr><br><DT><B>Maximum Likelihood Sequence Estimation for Asynchronous DataCommunications<p>Graduate Student: I. Sharfer<br>Professor: A. O. Hero<br></B><DD>We are developing techniques for maximum likelihood sequenceestimation for asynchronous multiple access communicationswith coherent spatial diversity using a receiver antennaarray. This project involves aspects of estimation theoryand lower bound analysis, iterative implementations ofmaximum likelihood (Viterbi and EM algorithms), multipleaccess communications, and antenna array processing (maximumlikelihood beamforming, direction finding, powerestimation). We have obtained a non-trivial extension ofthe Snyder-Georghiades sequqnce estimation algorithm to thearray receiever which includes spread spectrum modulationand the effects of Rayleigh fading. This has been achievedusing an iterative maximum likelihood algorithm based on ageneralization, called space alternating generalized EM(SAGE), of the expectation-maximization (EM) algorithm whichwas recently developed by Fessler-Hero for problems intomographic reconstruction. The resulting SAGE-typesequence estimation algorithm yields maximum likelihoodestimates which are of much lower complexity thanSnyder-Georghiades, involve no approximations, and areeasily generalizable to multipath and doppler shift commonin mobile radio communications.<hr><hr><br><DT><B>Research papers of Prof. D. L. Neuhoff andhis students</B> at the EECS Dept.of the University of Michigan can be accessed by anonymousFTP to <B>ftp.eecs.umich.edu</B> in the directory<B>/people/neuhoff</B>.<br>Or, from the EECS Dept. network, the path is<B>/n/ftp/f/people/neuhoff</B>.<br><hr><br><DT><B>Structured Vector Quantization and Asymptotic QuantizationTheory<p>Graduate Students: D. Hui, A. Balamesh, and D. Lyons<br>Professor: D. L. Neuhoff<br>Sponsors: National Science Foundation<br></B><DD>Vector quantization is increasingly being used as a lossydata compression technique for sources such as speech,images, and video. Practical vector quantizers "structure"their codebooks to simplify encoding and decoding. Forexample, block transform, CELP, tree-structured, two-stage,lattice, quadtree, product, pyramid, and finite-state vectorquantizers are common techniques, listed roughly indecreasing order of structure. Although structure generallyhas an adverse effect on rate/distortion performance, itpermits the use of quantizers with larger dimensions, whichusually results in much better performance for a givencomplexity. Until now there has been little theory toexplain the complexity performance tradeoff of structuredvector quantizers. <p>This project is developing new methods for analyzingstructured vector quantizers. One is an extension ofBennett's integral to vector quantizers. It shows how themean-squared error depends on the distribution and shape ofquantization cells. Another is an asymptotic formula forthe probability density of the quantization error. Thesenew methods have lead to the successful analysis of severalstructured vector quantizers, including tree-structured andtwo-stage quantizers. Still another is the analysis oftransform coders at very low rates.<p>The insight gained from this analysis has also led to a newform of two-stage quantization, called cell-conditionedmulti-stage VQ, that has the same low complexity advantagesof traditional multi-stage quantization, but asymptoticallysuffers no loss in performance relative to unstructuredquantization. It has also lead to new high performance, lowcomplexity methods for converting entropy coders intofixed-rate coders. <hr><br><DT><B>Model-Based Digital Image Halftoning<p>Professor: D. L. Neuhoff</B><DD>New model-based approaches to halftoning are beingdeveloped. They use well-known models of visual perceptionalong with models of printing that we have developed. Oneapproach minimizes the mean-squared error between theperceived intensity of the continuous-tone image and theperceived intensity of the printed halftoned image. Anotheris an adaptation of the well-known error diffusion method toinclude the printer model. Traditional approaches, forexample, ordered clustered dither, obtain robustness toprinter distortions, such as ink spreading, at the expenseof spatial resolution and the visibility of graininess. Incontrast, our new methods exploit the printer distortions toproduce higher quality images than would be obtained withRperfectS printers. Improvements due to model-basedhalftoning are expected to reduce the resolutionrequirements for laser printers used in high-qualityprinting (e.g., 400 dots/inch instead of 600). Model-basedhalftoning can be especially useful in transmission ofhigh-quality documents using high-fidelity, gray-scale imageencoders. In such cases, halftoning is performed at thereceiver, just before printing. Apart from codingefficiency, this approach permits the halftoner to be tunedto the individual printer, whose characteristics may varyconsiderably from those of other printers, for example,write-black vs. write-white laser printers.<hr><br><DT><B>Image Coding<p>Gracuate Students: M. Horowitz, M. Slyz<br>Professor: D. L. Neuhoff</B><DD>Image coding is the process of creating binary imagerepresentations with the dual goals of efficiency (as fewbits as possible in the representation) and accuracy (thereproduced images shall be as similar as possible to theoriginal). Two approaches are being pursued. The firstinvolves the use of a detailed model of the intermediatelevel human visual sensors to construct transform codes thathide quantization noise. The second involves the design oflossless image codes based on adaptive prediction, with newkinds of predictors and adaptation strategies. Theselossless image codes are intended for applications, such asmedical imaging, where an exact reproduction of the image isrequired. On the other hand, the first project is intendedfor more everyday applications where exact reproduction isnot necessary, but good quality and high efficiency areneeded. <hr><br><DT><B>Performance and Complexity of CDMA Networks withCoded-Modulation<p>Graduate Students: M. Klimesh, W. Sung<br>Professors: W. Stark and J. T. Coffey<br>Sponsors: National Science Foundation<br></B><DD>There are two parts of this research project. The first part dealswith decoding algorithms for worst case interference. Inthis work we have derived transmission and decodingstrategies that allows for maximum information transmissionor minimum error probability. These strategies allow thetransmitter to vary the transmission power pseudorandomly.The performance of a maximum likelihood decoding algorithmagainst the worst case jammer can be improved. Worst caseinterference is derived as well as the resulting⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?