Lee F. Holeva

2908 Mariposa Pl

Fort Wayne, IN 46818

260-452-7632

Examples of previous work:

A copy of my neural network paper from the Journal of Electronic Imaging may be found at:

Profile

Over twenty years of experience in product design, algorithm development, modeling, and system engineering. Research and product development interests include satellite imaging, machine vision, voice and image processing, neural networks, fuzzy inference and learning systems, artificial intelligence, optics and real-time embedded software. Have developed both hardware and software for one, two, and three dimensional signal acquisition and analysis with emphasis on embedded digital signal processing.

2006 to 2010: ITT/GS Fort Wayne

Staff Engineer: Member of the Modeling and Simulation Group.

Performed system engineering and testing activities including CRIS product data quality analysis during thermal-vacuum testing, system specification development using DOORS, and data quality analysis using measures such as SNR and NedT using tools developed in IDL.

Performed modeling and simulation for the CRIS, COMS, and ABI programs including both performance analysis and techniques for fringe count error detection. A patent on a new interferogram fringe count algorithm, based upon adaptive filtering, U.S. Patent number 7,764,383, is to be issued on July 27, 2010.

Performance modeling included the generation of simulated interferograms and Noise Equivalent change in Radiance, Nedn, predictions. Wrote Windows based C++ testing software for the ABI program as well as developing a Matlab based simulation using Kalman filtering of the line of sight navigation of the geosynchronous satellite.I also made a simulation of within channel crosstalk using the Framework Matlab based Model.

Analyzed and developed simulations for line of sight jitter for the HES program. Investigated algorithms for super resolution based upon the Richardson-Lucy algorithm.

Developed a mathematical model for the emissivity of a diffuse trap blackbody.

2003 to 2006: Kodak CGS/ITT Industries SSD, Rochester

Imaging Scientist: Responsible for image quality analysis

Setup and ran delta NIIRS studies to quantify image quality using tools developed in both Matlab and IDL. Dramatically increased productivity resulted from tools to aid image quality studies performing automatic image contrast adjustment using nonlinear optimization.

Developed tools to distinguish pattern noise from images having shot noise using wavelets. Models to simulate pattern noise were also developed. Image quality was also analyzed in term of Equivalent Noise Irradiance, NEI.

Developed techniques using Genetic algorithms to optimize tables for ADPCM coding. This supplemented traditional techniques of Lloyd-Max and Huffman encoding.

Developed a simulator of orbital smear. Starting from a Keplerian orbit, the position, terrain height, and attitude of the satellite may be varied and maps of smear versus time are produced.

1999 to 2003: Teradyne BTD, Deerfield, IL

Pattern Recognition/Neural Networks Engineer: Responsible for Pattern Recognition, signal processing, and neural network development for the testing of broadband testing systems. Prototypes were developed using a combination of Matlab and C++ code.

Invented hybrid RBF and MLP feed-forward neural networks for the pre-qualification of copper lines for DSL service. Another variant of the same hybrid neural nets, when used in pairs, was successfully shown to predict line capabilities, the maximum upstream and downstream speeds. Techniques from fuzzy logic as well as classical Levenberg-Marquardt were employed to train the neural networks. The nets were trained in Matlab with run-time code in C++.

Developed a neural network for the detection of bridge taps on copper telephone lines. Invented a method for the compensation of the switch testbus on line measurements. By employing open-loop adaptation the electrical effects of the switch was removed via deconvolution.

Setup SNMP software to access MIB information on the Lucent DSLAM.

Investigated the use of DOTNET (.NET) technologies, including C#, and ASP.NET, for the testing and configuration of home networks.

1998 to 1999: Vysis, Downers Grove, IL

Imaging Engineer: Responsible for C++ software development in the areas of image processing, pattern recognition, and artificial intelligence.

Invented a new algorithm based upon Fuzzy C-means clustering for the assignment of imaged two dimensional array of Fluorescent In-Situ Hybridization, FISH, spots to a two dimensional matrix.

Developed a new Radial Basis Function, RBF, neural network for chromosome classification. Using a two step process, first a basis layer in feature space is laid down using a variant of Fuzzy Learning Vector Quantization, FLVQ, followed by the adjustment of a layer of output weights by the Normalized Least Mean Square, NLMS, algorithm. Conditions for both the addition and the deletion of basis units are incorporated into the training algorithm.

1997 to 1998: 3Com Corporation, Skokie, IL

Senior Software Engineer: Responsible for DSP software development for full-duplex modem speakerphone on both Intel MMX and ADI platforms. Both individual software modules and complete subsystems were developed. Development tools such as the EZICE debugger and SoftIce have been used.

Developed critical modules for the USR Softmodem Speakerphone. These modules have included acoustic and line echo cancellers in Intel MMX as well as automatic gain control.

Designed and implemented the Hitman ADI 180x series modem speakerphone. The modem speakerphone was completely redesigned and optimized for the ADI 180x DSP.

1995 to 1997: Bell & Howell Postal Systems, Chicago, IL

Senior Software Engineer: Responsible for image processing algorithm development for a high-speed mail feeder. Tasks have included algorithm design, debug of code port to multiple C40 DSPs, and system integration. Program development proceeded from proof of concept on a UNIX box to the actual C40 DSP code debug. Special development tools such as a JTAG debugger were used here.

Developed algorithms for high-speed 4state bar code readers for flat mail. Barcodes were placed at any position and orientation on the magazines. Methods from mathematical morphology, the discrete wavelet transform, and maximum likelihood classification were used to detect barcodes. Fuzzy logic was utilized to classify each of the four bar types. Using a combination of special purpose image processing hardware and eight C40 DSPs, 4 state barcodes printed in black ink are decoded at rates exceeding 50,000 decodes per hour.

Developed an algorithm to detect address labels on flat mail. The problem was split into two parts, first a discrimination is made between foreground, which includes the label, and the background, which includes everything else. Then the foreground-background binary image is processing to detect rectangles, the assumed shape of the sought for label.

1990 to 1995: United Parcel Service Research & Development, Danbury, CT

Engineer: Responsible for algorithm and system development for both two and three dimensional machine vision inspection systems. Problems involved the machine classification of package shapes, the machine reading of coded labels, and optical character recognition, mostly in the context of moving conveyor belts.

Investigated the use of a feed-forward neural network to separate label characters from non-characters. A software simulation written in C demonstrated the successful utility of back-propagation trained networks applied to this problem (see paper #4).

Investigated both auto-focus and range measuring methods based upon an estimate of the amount of camera blur. For auto-focus, bracketing was combined with gradient descent to insure a final focused image. For dual camera ranging from blur special purpose optics were designed using Genii-Plus.

Developed a technique for the restoration of characters that are obscured by lines by means of a Hidden Markov Model.

Investigated the use of neural approaches, using Hopfield type networks and mean field annealing, for the binary segmentation of a gray scale label image.

Investigated the use of a Radial Basis Function (RBF) neural network to distinguish between handwritten and machine printed address labels. Estimates of local curvature and connected component length were manipulated to form feature vectors.

Developed a new feed-forward neural network architecture for the recognition of isolated characters via fuzzy structural features. The first layer of the new neural network architecture involved the creation of a map of local fuzzy features. A conventional multilayer perceptron type neural network may then used to recognize the character from a condensed version of the fuzzy feature map.

1986 to 1990: Allen-Bradley, Milwaukee, WI

Project Engineer: Responsible for both two and three dimensional vision system research.

Investigated algorithms, architecture, and applications for the NCR Gapp SIMD parallel processor

Developed the algorithms for edge location, sub-pixel interpolation, and camera calibration for the Configurable Vision Input Module (CVIM).

Investigated methods of performing three-dimensional gauging using binocular stereo. A trinocular system of three cameras was proposed to remedy problems inherent with stereo.

1984 to 1986: General Electric, Orlando, FL and Schenectady, NY

Engineer: Responsible for both two and three dimensional vision algorithm research. Problems involved the machine perception of rigid two and three-dimensional objects.

Developed algorithms for edge finding and enhancement of two-dimensional images. Applied gray scale morphology to range images.

1982 to 1984: University of Florida, Department of Electrical Engineering, Gainsville, FL

Research Assistant: Graduate research assistant under Prof. Alex Mystel.

A technique of robot path planning was proposed that decomposed a scene into a partition of weighted interlocking polygons.

1978 to 1982: RCA, Moorestown, NJ

Member of the Technical Staff: Performed both hardware and software design.

Investigated parallel SIMD architectures for digital signal processing. The primary application was for radar signal processing involving much use of discrete Fourier transforms.

Education

BSEE degree: University of Maryland, College Park, MD, 1978.

MSEE degree: Drexel University, Philadelphia, PA, 1982.

Post graduate work performed at the University of Florida, Gainsville, FL.

Additional graduate courses taken at Marquette University, University of Wisconsin, and the Illinois Institute of Technology.. Courses taken have included Kalman Filtering, Fourier Optics, Computational Geometry, and Signal compression and Vector Quantization.

Computer Language Proficiency

  1. Matlab
  2. IDL/ENVI
  3. C/C++/C# for MS Windows, UNIX/Linux/Xwindows, and TMS320C40 DSP.
  4. Assembly for Intel 586, TMS320C30, and TMS320C40 DSPs.
  5. Assembly for the Analog Devices 180x series DSP

Publications:

  1. Holeva, "Range Estimation from Camera Blur by Regularized Adaptive Identification", International Journal of Pattern Recognition and Artificial Intelligence, Vol. 8, No. 6, 1994.
  2. Holeva, "Range Estimation from Camera Blur by Regularized Adaptive Identification", SPIE Applications of AI: Machine Vision and Robotics, Orlando, FL, Vol. 1964, April 12-16, 1993.
  3. Holeva, "Learning How to Focus: Adaptive Control of an Autofocus Camera", SPIE High-Resolution Sensors and Hybrid Systems, San Jose, CA, Vol. 1656, pp. 383-393, Feb., 1992.
  4. Holeva, N. Kadaba, "An Image Filtering Module using Neural Networks for an Optical Character Recognizing System", 8th Intern. Conf. on Mathematical and Computer Modeling, College Park, MD, April 1-4, 1991.
  5. Holeva, "A Statistical Analysis of Subpixel Edge Finding", Intern. Robotics and Vision Conf., Detroit, MI, pp. 6-36 to 6-63, June 5-7, 1990.
  6. Holeva, "Foreground-Background Segmentation of OCR Labels by a Single Layer Recurrent Neural Network", SPIE Applications and Science of Artificial Neural Networks VI, April 17-21, 1995
  7. Holeva, Foreground-background segmentation of Optical character recognition labels by a single-layer recurrent neural network, Journal of Electronic Imaging 5(2), pp. 225-236, April, 1996.

Patents

  1. L. F. Holeva, Method and Apparatus for Passive Autoranging Using Relaxation, U. S. Patent No. 5365597, 11/15/94.
  2. L. F. Holeva, Method and Apparatus for Segmenting Foreground and Background Information for Optical Character Recognition of Labels Employing a Single Layer Recurrent neural Network, U. S. Patent No. 5710830, 1/20/98.
  3. L.F. Holeva, Automatic Gain Control with Adaptive Table Lookup, U.S. Patent No. 6,351,529, 2/26/2002.
  4. L.F. Holeva, Speedbinning by neural network, U.S. Patent No. 6,914,961, July 5, 2005.

Clearances

  1. current SSBI (black bar)