Rates of Convex Approximation in Non-Hilbert Spaces
M. J. Donahue, L. Gurvits, C. Darken, and E. Sontag
Constructive Approximation, 13, 187-220 (1997).
Abstract:
This paper deals with sparse approximations by means of convex
combinations of elements from a predetermined "basis" subset
S of a function space. Specifically, the focus is on the
rate at which the lowest achievable error can be reduced as
larger subsets of S are allowed when constructing an
approximant. The new results extend those given for Hilbert spaces by
Jones and Barron, including in particular a computationally attractive
incremental approximation scheme. Bounds are derived for broad
classes of Banach spaces; in particular, for Lp
spaces with 1<p<\infty, the O(n-1/2)
bounds of Barron and Jones are recovered when p=2.
One motivation for the questions studied here arises from the area of
"artificial neural networks," where the problem can be
stated in terms of the growth in the number of "neurons"
(the elements of S) needed in order to achieve a desired error
rate. The focus on non-Hilbert spaces is due to the desire to
understand approximation in the more "robust" (resistant to
exemplar noise) Lp, 1<=p<2 norms.
The techniques used borrow from results regarding moduli of smoothness in
functional analysis as well as from the theory of stochastic processes on
function spaces.
AMS classification: 41A25, 46B09, 68T05.
Title page.
Experimental Study of Magnetization Reversal Processes in
Nonsymmetric Spin Valve
V. S. Gornakov, V. I. Nikitenko, L. H. Bennett, H. J. Brown,
M. J. Donahue, W. F. Egelhoff, R. D. McMichael, A. J. Shapiro
Journal of Applied Physics, 81, 5215-5217 (1997).
Abstract:
We have investigated a nonsymmetric bottom giant magnetoresistance
spin valve with the structure Si/NiO/Co/Cu/Co/Ta, as well as single
ferromagnetic Co layers on antiferromagnetic NiO, with or without a
nonmagnetic Cu spacer. Magnetic hysteresis loops have been measured
by SQUID magnetometry, and magnetic domain structures have been imaged
using an advanced magneto-optical indicator film (MOIF) technique.
The MOIF technique demonstrated that the first stage of magnetization
reversal is characterized by nucleation of many microdomains. With
increasing reversed field, the domain walls move over small distances
(5-20 microns) until annihilation. The domain size was observed to
increase with the thickness of the Co layer. When an alternating
magnetic field was applied, the domain structure was dramatically
changed.
Title page.
Sparse Representations for Image Decomposition with Occlusions
Mike Donahue, Davi Geiger, Robert Hummel and Tyng-Luh Liu
Computer Vision and Pattern Recognition --- CVPR '96 (San
Fransisco, 7-12 (June 1996).
Abstract:
We study the problem of how to detect "interesting objects"
appearing in a given image, I. Our approach is to treat it as
a function approximation problem based on an over-redundant basis, and
also account for occlusions, where the basis superposition principle
is no longer valid. Since the basis (a library of image templates) is
over-redundant, there are infinitely many ways to decompose I.
We are motivated to select a sparse/compact representation of
I, and to account for occlusions and noise. We then study a
greedy and iterative "Weighted Lp Matching
Pursuit" strategy, with 0<p<1. We use an
Lp result to compute a solution and select the best
template at each stage of the pursuit.
Title page.
Image Recognition with Occlusions
Tyng-Luh Liu, Mike Donahue, Davi Geiger, and Robert Hummel
Proc. 4th European Conf. on Computer Vision --- ECCV '96
(London), 556-565 (April 1996).
Abstract:
We study the problem of how to detect "interesting objects" appearing
in a given image, I. Our approach is to treat it as a function
approximation problem based on an over-redundant basis. Since the
basis (a library of image templates) is over-redundant, there are
infinitely many ways to decompose I. To select the "best"
decomposition we first propose a global optimization procedure that
considers a concave cost function derived from a "weighted
Lp norm" with 0<p<=1. This concave
cost function selects as few coefficients as possible, producing a
sparse representation of the image and handling occlusions. However,
it contains multiple local minima. We identify all local minima so
that a global optimization is possible by visiting all of them.
Secondly, because the number of local minima grows exponentially with
the number of templates, we investigate a greedy "Lp
Matching Pursuit" strategy.
Title page.
Magneto-Optical Indicator Film Study of the Magnetization of a
Symmetric Spin Valve
V. I. Nikitenko, V. S. Gornakov,
L. M. Dedukh, Yu. P. Kabanov, A. F. Khapikov,
L. H. Bennett, P. J. Chen, R. D. McMichael, M. J. Donahue,
L. J. Swartzendruber, A. J. Shapiro, H. J. Brown,
W. F. Egelhoff, Jr.
IEEE Trans. Mag., 32, 4639-4641 (1996).
Abstract:
A magneto-optical indicator film (MOIF) technique is used for direct
experimental study of the magnetization reversal process in a
symmetric NiO/Co/Cu/NiFe/Cu/Co/NiO spin valve. It is shown for the
first time that the reversal of the free center layer proceeds by
nonuniform magnetization rotation. The observed switching mechanism
is presumed to be associated with the influence of the nonuniform
magnetostatic field that follows from surface roughness and the
polycrystalline structure of the magnetic layers.
Title page.
Complementary Imaging of Granular Co-Ag Films with Magneto-Optical
Indicator Film Technique and Magnetic Force Microscopy
M. J. Donahue, L. H. Bennett, R. D. McMichael, L. J. Swartzendruber,
A. J. Shapiro,
V. I. Nikitenko, V. S. Gornakov,
L. M. Dedukh, A. F. Khapikov,
V. N. Matveev, V. I. Levashov
Journal of Applied Physics, 79, 5315-5317 (1996).
Abstract:
A Magneto-optical indicator film (MOIF) technique and magnetic force
microscopy were used for visualization and direct real-time
experimental study of the magnetization processes of magnetic
Co90Ag10 granular films. It is shown that the
magnetization reversal of the as-deposited films follows a specific
two-step course. The first stage is characterized by gradual spin
rotation to large angles up to 90° without domain formation.
Further magnetization reversal proceeds by the nucleation and motion
of zigzag-shaped domain walls. The dendritic structure of the domain
walls was observed using both techniques. Also tracks of magnetic
inhomogeneities remaining behind moving zigzag-shaped domain walls
were revealed by MOIF.
PACS: 75.60.Ch, 61.16.Ch, 75.60.Lr
Title page.
Artifacts in Magnetic Resonance Imaging from Metals
L.H. Bennett, P.S. Wang, and M.J. Donahue
Journal of Applied Physics, 79, (1996).
Abstract:
Metallic biomedical implants, such as aneurysm clips, endoprostheses,
and internal orthopedic devices give rise to artifacts in the magnetic
resonance image (MRI) of patients. Such artifacts impair the
information contained in the image in precisely the region of most
interest, namely near the metallic device. Ferromagnetic materials
are contraindicated because of the hazards associated with their
movement during the MRI procedure. In less-magnetic metals, it has
been suggested that the extent of the artifact is related to the
magnetic susceptibility of the the metal, but no systematic data
appear to be available. When the susceptibility is sufficiently
small, an additional artifact due to electrical conductivity is
observed. We present an initial systematic study of MRI artifacts
produced by two low susceptibility metals, titanium (relative
permeability µr approx. 1.0002) and copper
(µr approx. 0.99998), including experimental,
theoretical, and computer simulation results.
PACS: 87.59.Pw, 75.20.En
Title page,
Presentation page.
On the Use of Level Curves in Image-Analysis
M.J. Donahue and S.I. Rokhlin
CVGIP-Image Understanding, 57, 185-203 (1993).
Abstract:
A digitized image is viewed as a surface over the xy-plane. The level
curves of this surface provide information about edge directions and
feature locations. This paper presents algorithms for the extraction
of tangent directions and curvatures of these level curves. The
tangent direction is determined by a least-squares minimization over
the surface normals (calculated for each 2 x 2 pixel neighborhood) in
an averaging window. The curvature calculation, unlike most previous
work on this topic, does not require a parameterized curve, but works
instead directly on the tangents across adjacent level curves. The
curvature is found by fitting concentric circles to the tangent
directions via least-squares minimization. The stability of these
algorithms with respect to noise is studied via controlled tests on
computer generated data corrupted by simulated noise. Examples on real
images are given which show application of these algorithms for
directional enhancement, and feature point detection.
Title page.
A Method for Automatic Inspection of Printed-Circuit Boards
A.P. Sprague, M.J. Donahue, and S.I. Rokhlin
CVGIP-Image Understanding, 54, 401-415 (1991).
Abstract:
An algorithm for automatic inspection of printed circuits is discussed
which includes elements of the reference-comparison and the
design-rule methods. First a directed graph, called the segment graph,
is created from the image of the PCB. Construction of this graph
involves comparison of adjacent rows, and hence each pixel is examined
only twice in this construction. Feature points, including wire
intersections, wire ends, and circuit defects, which uniquely
represent the PCB image are extracted from the segment graph. In this
way a very compact representation of the image is extracted. The
feature points of the board being tested are then matched against the
feature points from a reference board. Simultaneous with feature point
extraction, the design-rule checking algorithm performs wire width and
spacing inspection, and marks the unsatisfactory regions. Results of
the algorithm implementation are demonstrated.
Title page.