 |
Description  |
|
|
The invention relates to a method and apparatus for electronic
stabilization of image sequences using distal image processing to remove
an unwanted component of frame to frame motion.
BACKGROUND OF THE INVENTION
The need for image stabilization arises in many applications including the
movie industry which must remove unwanted jitter between successive frames
of source video, television news cameramen who must stabilize video from
hand held cameras in newscasts produced in the field, video from
surveillance cameras molted on swaying or rotating platforms, or on moving
vehicles, which must be stabilized prior to computer analysis, or prior to
display to human observers, and video from moving vehicles which must be
stabilized prior to image compression or presentation to a remote operator
(teleoperation) or to computer vision systems for automatic driving.
Mechanically stabilized platforms of various types are used in surveillance
to compensate for imager, not image, motion. In FIG. 1, an imager 10 is
mounted on a mechanically stabilized platform 12. The output I.sub.in (t)
on lead 14 of the imager 10 is displayed on a monitor 16. The platform 12
typically uses gyroscopes to sense platform rotation, and motors to
compensate for that rotation. A user can guide the imager direction of
gaze (pan and tilt) and zoom via an electronic control sisal 18 and motor
drives in the platform 12.
Electronic stabilization with imager motion sensors can be used to
compensate for imager motion but not image motion. In FIG. 2, an imager
10, mounted on a mechanically stabilized platform 12, has an output
I.sub.in (t) on lead 14. Residual motion of the platform 12 is measured
using sensors 20. Sensed displacements d(t) on lead 22 are converted to
transformation parameters p(t) on lead 24 by a transform module 26.
Parameters p(t) 24 are used by image warp deuce 28 to produce a stabilized
output image I.sub.out (t) 30, in which imager motion has been
compensated, for display on monitor 16.
In FIG. 3 a system for electronic stabilization with digital processing to
sense image motion is shown. The system includes imager 10 having an
output I.sub.in (t) 14 which is stored in an image frame store 32 to hold
the image until appropriate warp parameters have been computed and
transmitted to image warp device 28 for stabilization. (This frame store
is not needed if the warp performed at one frame time is based on
parameters computed at the previous frame time.) In order to reduce
memory, a set of features may be extracted from the source video by module
34. A second image frame store 36 is provided to hold the features
extracted from the previous image so that it can be compared with the
present image. The features f(t) on lead 38 extracted at time t are
compared with features f(t-1) on lead 40 extracted at time (t-1) in
displacement estimator 42 that uses digital image processing to determine
image-to-image motion to produce displacements d(t) on lead 44. One
commercially available camera uses an array of 36 pixels as the features
compared between frames. Sensed displacements d(t) on lead 44 are used by
image warp device 28 to produce a stabilized output image I.sub.out (t)
30, in which image motion has been compensated, for display on monitor 16.
Such systems known to the inventor can not compensate zoom, rotation,
parallax and/or lens distortion.
In FIG. 4 an electronic target tracking system uses correlation to locate a
target within an imager's field of view, then steers the imager to center
the target, and, at least roughly, stabilize the target pattern. The
system includes imager 10 having an output image I.sub.in (t) 14 which is
compared in a correlation module 46 to match a reference pattern stored
therein with the image 14. The reference pattern is selected from a prior
frame in the video sequence module 48, stored in memory 50. The reference
pattern may represent a stationary object in the scene or a moving object.
The difference between the image I.sub.in (t) and the reference pattern
provides displacement information to transform module 26 which converts
this information to produce a signal to stabilize platform 12.
In FIG. 5 a system for electronic target tracking using change detection is
shown. The difference between the current image I.sub.in (t) and a
previous image I.sub.n (t-1) in frame store 60 is determined by subtractor
62 and regions of significant change are located. The location information
x(t) provided by module 64, is used to redirect the imager 10 to maintain
the image at a particular point or on a particular track in the displayed
image. This approach may be preferred to a pattern based approach when
targets are too small to be detected based on pattern match, but are
moving so that detection can be based on this motion. Existing systems
require that motion of the background scene be small so that target motion
can be detected.
Burt et al. in the Proceedings Of The Workshop On Visual Motion, Irvine,
Calif., Mar. 20-22, 1989, pages 1-12, have disclosed procedures for
obtaining precise image alignment through iterative refinement. Such
procedures have been used for such applications as terrain shape recovery
and moving target detection. It would be desirable to have a method and
system for electronically stabilizing images which can compensate zoom,
rotation, parallax and/or lens distortion while having improved accuracy.
It would also be desirable to extend the functionality of an image
stabilization system to include derived image sequences with reduced noise
or highlighted change by making use of information generated as a
byproduct of the improved stabilization process.
SUMMARY OF THE INVENTION
The invention is an apparatus and method for electronic stabilization of an
image. The input may be any sequence of image frames from an image source,
such as a video camera, an IR or X-ray imager, radar, or from a storage
medium such as computer disk memory, video tape or a computer graphics
generator. The output of the system is a modified image sequence in which
unwanted components of image motion have been reduced or removed, or in
which new motion of a desired form has been introduced. The output may
also include information derived from the input sequence, such as
estimated frame-to-frame displacement vectors or motion flow fields, and
derived image sequences with reduced noise or highlighted change.
The invention is a method for stabilizing an image relative to a reference
comprising the steps of (a) warping the reference using an initial
estimate of the displacement between the image and the reference; (b)
comparing the warped reference with the image to obtain a residual
displacement; (c) adding the residual displacement to the initial estimate
to obtain a revised estimate of the displacement between the image and the
reference; (d) warping the reference using the revised estimate of the
displacement; (e) repeating steps (b), (c) and (d) K times, thereby
forming a final estimate of the displacement d(t; K); (f) forming output
transformation parameters p(t) from the final estimate of the displacement
d(t; K); and (g) warping the image using output transformation parameters
p(t) to form an output image.
The invention is also apparatus for stabilizing an image relative to a
reference comprising means for warping the reference in response an
estimate of the displacement between the reference and the image; means
for comparing the warped reference with the image to obtain a residual
displacement; means for adding the residual displacement to the initial
estimate to obtain a revised estimate of the displacement between the
image and the reference; means for warping the reference using the revised
estimate of the displacement; means for repeating steps (b), (c) and (d) K
times, thereby forming a final estimate of the displacement d(t; K); means
for forming output transformation parameters p(t) from the final estimate
of the displacement d(t; K); and means for warping the image Iin(t) using
output transformation parameters p(t) to form an output image.
BRIEF DESCRIPTION OF THE DRAWING
Elements common to different Figures are given the same numerical
identification in each of the Figures.
FIG. 1 is a schematic diagram of a prior art mechanically stabilized
platform;
FIG. 2 is a schematic diagram of a prior art electronic stabilization
system with imager motion sensors;
FIG. 3 is a schematic diagram of a prior art electronic stabilization
system with digital processing;
FIG. 4 is a schematic diagram of a prior art system for electronic target
tracking;
FIG. 5 is a schematic diagram of a prior art system for electronic target
tracking via change detection;
FIG. 6 is a schematic diagram of an electronic stabilization system of the
invention;
FIG. 7 is a schematic diagram of another electronic stabilization system of
the invention;
FIG. 8 is a schematic diagram of a system for motion adaptive
frame-to-frame processing;
FIG. 9 is a schematic diagram of another stabilization system using a
separate reference pattern, R;
FIG. 10 is a schematic diagram of another stabilization system using two
image sequences;
FIG. 11 is a schematic diagram of a system for locating a target;
FIGS. 12 and 13 illustrate methods of obtaining the transformation
parameters;
FIG. 14 illustrates a method for forming the Gaussian and Laplacian
pyramids;
FIG. 15 is a schematic diagram of a PYR-1 circuit;
FIG. 16 is a schematic diagram of an alternative embodiment of the FIG. 6
electronic stabilization system.
DETAILED DESCRIPTION
The invention is a system and method for electronically stabilizing an
image produced by an electronic imaging device. The input may be any
sequence of image frames I.sub.in (t) from an image source, such as a
video camera, an IR or X-ray imager, radar, or from a storage medium such
as computer disk memory, video tape or a computer graphics generator. The
invention can also be used with images from multiple sources when these
must be stabilized with respect to one another, with one image being
stabilized and the other image providing the reference. The invention uses
a feedback loop and second image warp stage to achieve precise image
alignment as part of the displacement estimation process. The output of
the stabilization system is a modified image sequence, I.sub.out (t), in
which unwanted components of image motion have been reduced or removed, or
in which new motion of a desired form has been introduced. The output may
also include information derived from the input sequence, such as
estimated frame-to-frame displacement vectors or motion flow fields, and
derived image sequences with reduced noise or highlighted change.
In FIG. 6 an electronic stabilization system 100 of the invention comprises
an imager 10 having an output I.sub.in (t), an image frame store 32 for
storing the image I.sub.in (t), and a second frame store 36 for storing a
previous image I.sub.in (t-1). The contents I.sub.in (t-1) in frame store
36 is warped in prior frame image warper 102 using an initial estimate
d(t; 0) of the displacement provided by a feedback loop 104, shown in the
dotted box, to provide a warped prior image I.sub.in (t-1; k-1). The
warped prior image I.sub.in (t-1; k-1) is compared in means 106 for
estimating displacement with the original image I.sub.in (t) to obtain an
estimate of the residual displacement .DELTA.(t; k) which provides an
input to the feedback loop 104 from the displacement estimator 106. The
residual displacement .DELTA.(t; k) is then added to the previous estimate
of the displacement in adder 108 and held in memory 110 to produce a
refined estimate of the displacement d(t; k)=d(t; k 1)+.DELTA.(t; k).
These steps are repeated K times until the residual displacement is less
than a given value or, alternatively, a fixed number of times selected by
an operator or other means, thereby forming a final estimate of the
displacement. Terminal 112 is connected to a means 114 for converting the
final estimate of the displacement d(t; k) to output transformation
parameters p(t). This includes any image change which can be modelled,
including zoom, most parallax, translation, rotation, dilation and lens
distortion. The output of the means 114 is connected to the present image
warper 116. Present image warper 116 warps the present image I.sub.in (t)
to form I.sub.out (t) at terminal 118. The refined estimate of the
displacement d(t; K) after K iterations in the feedback loop 104 is also
available at output terminal 120. The estimates of the
image-to-reference/image displacement can be based on displacement
measures computed over only parts of the image domain, e.g., those
containing targets patterns.
In the method of the invention, iteration k begins with a prior estimate
d(t; k-1) of the displacement from I.sub.in (t-1) to I.sub.in (t). Image
I.sub.in (t-1) is warped by d(t; k-1) to form the warped image I.sub.warp
(t-1; k-1). The displacement estimation process is applied to the current
I.sub.in (t) and the warped image I.sub.warp (t-1; k-1) to obtain an
estimate of the residual displacement Dd(t; k). The residual displacement
.DELTA.d(t; k) is then added to the prior estimate of displacement to
obtained a refined estimate of displacement d(t; k)=d(t; k-1)+.DELTA.d(t;
k). The image warp step is then repeated to form I.sub.warp (t-1; k) that
is more precisely aligned with I.sub.in (t). These steps are repeated K
times to obtain a final estimate d(t; K) of the displacement from frame
I.sub.in (t-1) to frame I.sub.in (t). The prior estimate used in the first
iteration of this process, d(t; 0), may be based on the final estimation
obtained at the previous frame time, e.g., d(t; 0)=d(t-1; K). The cycle at
time t is then completed by transferring a copy of I.sub.in (t) from frame
store 32 to frame store 36.
In this embodiment, the prior, or reference, frame I.sub.in (t-1) is warped
into registration with the current frame, I.sub.in (t) to provide the
displacement information. Alternatively a reference from the same or
another source can be used to provide the displacement information to warp
the current frame into registration. Typically the initial estimate of
displacement d(t; 0) used at time t is taken to be equal to the final
estimate d(t-1; K) at time t-1. If acceleration from frame to frame is
small it may only be necessary to perform one iteration of the estimation
process per frame time, K=1, to achieve successive refinement.
While the proposed stabilization method makes use of two distinct warp
steps, in practice these can often be performed by the same warp hardware
device by multiplexing the warper inputs and outputs. For example, the
output from frame store 36 and loop 104 are inputs and warper output to
the estimator 106 is an output. Alternative inputs are the present image
and the transformation parameters and the warper output is then the output
image.
The images I.sub.in (t) and I.sub.warp (t-1; k-1) may be filtered or
otherwise processed prior to displacement estimation as shown in FIG. 7.
In a preferred implementation, a Laplacian pyramid is constructed for the
current input image, I.sub.in (t) in pyramid processor 130 and stored in
memory 32. The Laplacian pyramid for I.sub.warp (t-1; k-1) is regenerated
after the warp step on each iteration, k in the pyramid processor 134.
Then successive refinement steps of the estimation process may base the
estimate of residual displacement on progressively higher resolution
pyramid levels.
In FIG. 8 a system 150 for motion adaptive frame-to-frame processing as
part of an image stabilization system of the invention includes means 152
for summing the present image I.sub.in (t) and the warped image I.sub.warp
(t-1; k-1) pixel by pixel (temporal lowpass), to obtain a composite
A(t)=I.sub.in (t)+I.sub.warp (t-1; k-1)
with reduced image noise. The system 150 also includes means 154 for
subtracting (temporal high-pass) the present and the warped images pixel
by pixel to form a difference image
B(t)=I.sub.in (t)-I.sub.warp (t-1; k-1)
that reveals change or motion in the scene (moving target detection).
The sum and difference images may be provided directly as outputs of the
stabilization system in addition to, or in place of, the stabilized output
I.sub.out (t). The derived image sequences may be stabilized by warping in
the same way that an image is stabilized with reference to FIG. 6.
More than two images may be aligned either for noise reduction (temporal
low pass filter) or change detection (temporal high pass). As options,
noise reduced images may be further enhanced, e.g., through sharpening or
noise coring, or the change image can be further processed, e.g., to form
a change energy image (or pyramid). It should be noted that methods for
image alignment and temporal filtering are known. The invention is thus
also the incorporation of frame-to-frame processing of images that are
aligned as a step in image stabilization. In particular the generation of
noise reduced or change enhanced image sequences through temporal low- and
high-pass filtering, respectively, of the aligned input images and the
stabilization of these derived sequences prior to output from the system.
In FIG. 9 a system 160 for aligning each image of the source sequence with
a separate reference pattern R rather than with the preceding image. The
reference pattern may be obtained from various sources. Typically, the
reference pattern R is selected by selector 162 under user control from an
earlier frame of the source video or it may simply be a previous image or
piece of a previous image in the source video (e.g., every n.sup.th frame
may be selected and used as the reference to align the following n-1
frames), or it may be a selected from a set of stored target patterns. The
pattern R, stored in frame store 36, is warped in warper 102, producing an
output R.sub.warp (k-1) in response to an estimate of the displacement of
the pattern R between the reference and the present image Iin(t). The
output of the system is the present image I.sub.in (t) warped to hold the
region matched by the reference pattern stationary, or to cause that
region to move along an arbitrary desired path. Alternatively, rather than
search for a single target pattern, the system may search for a set of
component patterns, e.g., using hierarchical structured search.
In FIG. 10 a system 180 for stabilizing images from different sources,
e.g., IR and visible imagers, by registration of a particular image
I.sub.1n (t) generated in imager 10 and stored in frame store 32 to a
second image sequence I.sub.2n (t) from an image source 182 requires
special displacement estimation procedures. The second image sequence
I.sub.2n (t) is stored in frame store 184. The images in one source are
aligned with the images from the other source; i.e., images from one
source serve as the reference patterns used to stabilize the images from
the other source. Warped image I.sub.warp (t; k-1) is compared with image
I.sub.1n (t) in means 106 for estimating differences to determine the
present difference d(t; k) which is then added to the previous estimate
d(t; k-1) to provide the new estimate to warper 102. The final estimate of
the differences is used by means 114 for generating transform parameters
to provide the necessary information to warper 116 for stabilizing
I.sub.in (t) to produce image I.sub.out (t) at terminal 118. Image
I.sub.2n (t) is also available as a system output at terminal 186.
Various processes can be used to detect the location of target objects
within the imager's field of view. These include correlation match (e.g.,
registration to a reference pattern), the detection of a region of
activity in the motion flow field, and the detection of a region of
activity in the temporal change image. Prior art systems use electronic
target detection to steer the imager. These rely on correlation pattern
match or frame to frame change but do not make use of image motion flow or
of frame to frame change after electronic alignment. As a result, such
systems can only use frame to frame change to detect a moving target if
the imager is tracking the background and the background image is
essentially stationary.
In the method of the invention, the background is electronically aligned,
then the target is detected, then the location of the target directs the
imager and/or electronic stabilization process to track and stabilize the
target, not the background, in the output image sequence. The invention
uses derived change or motion information to detect targets moving
relative to the background, and to estimate both target motion and
background motion and electronically detected target location and
background motion information to control both imager steering and output
image sequence stabilization.
In FIG. 11 a system 200 icludes high pass filter means 202 which takes the
difference I.sub.in (t)-I.sub.warp (t-1) and means 204 for determining the
location of a target; i.e. any rapidly changing object in the images. The
output of means 204 provides positional information which is fed to the
means 106 and to a second means 206 for generating transform parameters.
The output of means 206 is used to control imager platform 208 to steer
the imager, e.g., in order to track or center the target, or to
electronically stabilize or center the target. The input to means 106
provides offset motion to the image which can be used, for example, to
center a rapidly moving object in the output image. More generally, the
target location information can be used to both steer the imager and
electronically compensate for residual motion. In this case warp
parameters are based both on flame-to-frame displacement d(t; K) and on
target location parameters x(t). The combination of filter means 202 and
locator means 204 can also accept filtered differences from other pairs of
images which are spatially and temporally aligned to improve target
detection and reduce noise.
The function of an image stabilization system is to determine motion of the
input sequence and replace observed motion with a desired motion in the
output sequence. There are various transformations that can be used in
means 114 to convert observed displacements d(t) to required warp
parameters p(t). Here we describe three such methods. In FIG. 12, let
x.sub.in (t) be the location of the t.sup.th input frame. In general this
is a vector quantity including x, y, .THETA., scale (or higher order
parameters in the case of model based stabilization, described below). Let
.DELTA.x.sub.in (t)=d(t; K) be the observed frame-to-frame displacement.
Let x.sub.desired (t) be the location of the t.sup.th image on some
desired output path, P. The means 114 determines the displacement
.DELTA.x.sub.shift (t) required to shift images from observed input
locations to desired output locations. The transform module also
determines the desired path P based on observed input displacements. The
following are methods for obtaining P (x.sub.desired) from the observed
.DELTA.x.sub.in.
In a "low-pass" method for transforming displacement estimates, the output
sequence of positions is obtained by simply smoothing the input sequence
of positions:
.DELTA.x.sub.shift (t)=.DELTA.x.sub.in (t)-w*.DELTA.x.sub.in (t).
Here w is typically an IIR low pass filter. Different filters may be
applied to obtain different parameters of motion. For example rotation
.THETA. may be heavily smoothed, while translation (x, y) is moderately
smoothed, and scale is not smoothed at all. In practice non-linear low
pass filters may be used. For example there may be a limit placed on how
large the displacement x.sub.shift (t) is allowed to get. The invention is
also the motion damping by applying a low pass filter to the sequence of
input positions to determine the desired output positions.
In a piece-wise continuous method fopr transforming displacement estimates
as shown in FIG. 13, the output path is defined as a sequence of jointed
linear (or other smooth) segments. These segments are selected to follow
the expected trend in input image positions. All output frames may be
shifted to a line segment A until time t.sub.1 is reached at which this
shift exceeds a predefined limit. Then a new segment is initiated that
will tend to bring the output images back into the center of the output
display. An estimate of velocity at t.sub.1 is based, for example, on an
average of the observed input displacement for a period of time preceding
t.sub.1. Alternatively, the desired path may be specified externally. This
velocity and the position of frame I.sub.in (t.sub.1) define a path B that
predicts positions of subsequent input frames. The output path P is then
made to follow a transition segment C that is constructed from the end of
segment A, at time t.sub.1 -1, to segment B at t.sub.2. Then P follows B
until such time as the required shift again exceeds the predefined limit.
This method removes high temporal frequency motions of the input sequence
while allowing the output to follow trends in the input motion that may
change abruptly in direction.
When a target tracking process is used in combination with a image
alignment process as shown in FIG. 11, the detected target position can be
used to determine output path P while the image alignment process is used
to shift images to that path. The two-process approach may be needed when
targets are small or move somewhat erratically in the scene. Piece wise
smooth image stabilization ensures that unwanted high frequency motion of
the background scene is removed in the output. Target tracking is used to
define a path that keeps the target generally centered within the output
display.
The target pattern may be a piece of the background scene designated by the
operator, or a moving object in the scene designated by the observer or
detected by the motion detection module. In the case of a pattern
designated by the observer the system stores a pattern to be used as a
reference in subsequent tracking. This reference may be automatically
updated at regular intervals, e.g., every n.sup.th frame. Once the
location that provides a best match to the reference pattern is found in
the scene, the pattern at that location is stored as a replacement for the
reference. In this way the system can adapt to gradual changes in the
target pattern. If the detected location of the target pattern were used
directly to stabilize the output sequence there would tend to be a slight
jump (discontinuity in output path, P) each time the reference pattern is
updated. This jump is avoided by using observed target pattern position to
define segments in the path, P, while the background precise alignment
process defines shifts x(t) and p(t) in FIG. 11. The invention is also the
use of separate background and target tracking processes in stabilization.
The background process provides piece wise smooth stabilization while the
target tracking process determines segments of the output motion path in
order to move the target pattern smoothly to a desired position in the
output display.
Existing mechanical stabilization devices stabilize two or three axes of
imager rotation. Existing electronic devices compensate for image
translation in the imager field of view. An important advantage of the
electronic approach proposed here is that it can be used to compensate for
components of motion that cannot be compensated through imager control.
For example, electronic warp can compensate for lens distortions as
patterns move over the imager field of view. More importantly, electronic
stabilization can be use to compensate for parallax motion as the imager
moves relative to a surface in the scene. Stabilization is based on a
mathematical model of that motion and parameters estimated through the
precise alignment method outlined above. For example a quadratic warp can
compensate for motion of a planar surface that may be tilted relative to
the imager direction of gaze.
Electronic stabilization may be used to compensate for the motion of
surfaces that exist in the scene (e.g., a road surface in front of an
autonomous vehicle) or of surfaces that are define relative to the imager
(e.g., a vertical surface at a predefined distance from a moving vehicle).
The former method provides a means for estimating the motions and
distances to surfaces in the scene as well as stabilizing those surfaces,
while the latter method provides a means for detecting objects as they
cross an "invisible barrier" at a prescribed distance from the imager.
(Detection can be based on residual motion model based stabilization of a
surface in front of a moving vehicle followed by temporal low pass filter
results in motion blur for all objects in the scene that are not near that
surface. Objects will appear sharp as they pass through the "invisible
barrier".)
Model based stabilization can be used to stabilize a surface in the scene,
such as that of the road itself. In this case the motion of the surface
relative to the imager must be determined. This can be done using known
model based displacement estimation procedures. These techniques have been
proposed in the past as a means for determining surface shape and
distance. Here we propose the use of these techniques in the stabilization
of image sequence.
Imaging means 10 may be comprised of a structure for receiving radiation
reflected from objects within a field-of-view which are illuminated by an
external radiator. Alternatively, imager means 10 may be comprised of a
structure that includes means for illuminating objects within its
field-of-view with radiation and means for receiving reflected echoes from
such objects (these echoes also may provide object-distance information).
Further, imaging means 10 may be responsive to radiation of any given
wavelength portion of electromagnetic, ultrasonic and/or any other type of
wave-energy spectrum. An analog-to-digital (A/D) converter converts the
image data in each successive frame to digital form for processing by a
digital processor
The image warpets 102 and 116 are preferably formed from a TMC 2302 Image
Manipulation Sequencer and a TMC 2246 Image Filter manufactured by TRW,
Inc. The Image Manipulation Sequencer is used to generate an address in
the source image given an address in the output image. The Image Filter
interpolates the output pixel value given input pixels in the neighborhood
of the specified input image address. The input image is stored in
standard memory that has been segmented to allow simultaneous read of all
the neighboring pixels needed by the interpolator.
The transform unit converts observed image displacements and externally
spacified control information. The transform module may be a general
purpose computing device, such as the TMS 320 C 30 digital signal
processing chip manufactured by Texas Instruments, Inc., Dallas, Texas,
with some memory. The transform module may include several different
stabilization control modes, each specified by computer code stored in its
memory. For example, these may include damped motion stabilization,
piece-wise linear stabilization, and stabilization combined with target
centering. The mode used at any moment in time is normally specified by an
external input to the system. Other inputs to the transform module include
frame to frame displacements, computed by the displacement estimator, and
target locations, computed by the locator module. Each frame time the
transform module first determines the desired output motion path, then the
required displacement to move an input frame to that path. The
displacement parameters are provided to the warper as the output of the
system. Inputs include frame to frame displacments information from the
displacement estimator and external mode or other control specification.
Image analysis by decomposition of an image where a comparatively
high-resolution image having a first number of pixels is processed to
derive a wide field-of-view, low resolution image having second number of
pixels smaller than the first number has been disclosed by Burt in
Multiresolution Image Processing And Analysis, volume 16, pages 20-51,
1981 and Anderson et al in U.S. Pat. No. 4,692,806, incorporated herein by
reference for its teachings on image decomp | | |
