Contributed Session: Image Analysis
Efficient Estimation of Process Parameters from Image Sensor Data
Daniel C. Chin
A very efficient stochastic optimization technique, simultaneous perturbation stochastic approximation (SPSA) introduced by J. C. Spall, has a wide range of applications including image analysis, experimental design, optimal control, and parameter identification and estimation. This paper presents an SPSA application for the problem of extracting the evolved model structure from global magnetospheric images. Similar problems arise, for example, in buried waste recovery and sensing and process fault isolation through image data. The parameter estimation in this problem involves complicated modeling, expensive line of sight integration, and an iterative optimization procedure. A simulation study with noisy data shows that SPSA- derived parameter estimates approach the true values and the convergence speed increases dramatically due to a decrease from 300-400 function evaluations (using a Powell algorithm) to 30-40 function evaluations. Most importantly, the procedure recovers the true model (which generates the images) while saving computational and personnel costs, and it will permit a greater degree of flexibility in the functional design of the data gathering mission.
[Daniel C. Chin, Applied Physics Laboratory, Johns Hopkins Univ., Laurel, MD 20723-6099 USA; email@example.com ]
A Markovian Approach to Microstructure Modeling of Plasma-Sprayed Coatings
Robert E. Derr
Images of the microstructure of materials can be simulated using a germ-grain model to tesselate the plane or a Monte-Carlo technique to iteratively build the cells. While these methods work reasonably well for some multi-phase microstructures, they do not properly simulate the two-phase microstructure resulting from a plasma-sprayed coating.
We propose applying a nearest-neighbor marked Markov point process to model and simulate the pores of the microstructure of plasma-spray coated materials. Issues that arise involve the shape and size of the marks (pores), the concentration of the pores, and the further modeling of the `splats' resulting from the plasma-spray process given the location of the pores.
The methods are based on work by Baddeley and Moller (Int. Statis. Rev. 57 (1989):89--121).
[Robert E. Derr, Dept. of Statistics, Univ. of North Carolina, Chapel Hill, NC 27599-3260 USA; firstname.lastname@example.org ]
Development of a Particle-in-Oil Standard for the National Fluid Power Industry
The National Fluid Power Industry has requested the development of a standard reference material (SRM 2806) composed of a mineral dust suspended in hydraulic fluid. SRM 2806 is being certified for particle size distribution per fluid volume, to be used in conjunction with the NFPA "Hydraulic Fluid Power - NIST Traceable Calibration Method for Liquid Automatic Particle Counters." SRM 2806, dust-in-oil, was produced by suspending ISO medium dust in hydraulic oil at a nominal concentration of 2.8 mg/L. The material samples were tested for batch homogeneity using commercial optical particle counters at the manufacturer and at NIST. The number distribution of the SRM is being certified by electron microscopy coupled with image processing. Image processing software, developed at NIST, uses thresholding and "blobbing" to derive the particle number size distribution from archived images. The measurement techniques developed to certify this material, the image processing and the analytical results will be discussed.
[R.A. Fletcher, Surface & Microanalysis Div., NIST, Gaithersburg, MD 20899 USA; email@example.com ]
Adaptive Smoothing of Images with Local Weighted Regression
Mark S. Levenson
David S. Bright
We present a weighting scheme for local weighted regression designed to achieve two goals: (1) Reduce noise within image regions of smoothly varying intensities; and (2) Maintain sharp boundaries between image regions. The procedure parallels that of Cleveland's lowess procedure. An iteration of the procedure consists of the weighted least-squares fitting of polynomials to local neighborhoods of each pixel. The weight of a pixel in a fit is based on a gradient estimate from the previous iteration of the pixel relative to the central pixel. A pixel with a large gradient is downweighted in the regression fit. The weighting function has a monotone property in which weights cannot increase along a direction from the central pixel. The degree of downweighting is adaptive to the scale of the image using quantiles of local gradients.
[Mark S. Levenson, Statisical Engineering Div., NIST, Gaithersburg, MD 20899 USA; firstname.lastname@example.org ]
Date created: 6/5/2001