DESCRIPTION
STEREOSCOPIC IMAGE PROCESSING DEVICE, METHOD, RECORDING
MEDIUM AND STEREOSCOPIC IMAGING APPARATUS
Technical Field
The presently disclosed subject matter relates to a stereoscopic image processing
device, a stereoscopic image processing method, a recording medium and a stereoscopic
imaging apparatus. More particularly, the presently disclosed subject matter relates to
an art geometrically correcting variations in angles of view between a plurality of images
due to placement attitudes of a plurality of imaging units and variations in zoom factor
and the like.
Background Art
A camera including a plurality of imaging units in one camera apparatus (a so-
called compound eye camera) can image the identical subject from a plurality of
viewpoints. Thus, the compound eye camera can be used for acquiring three-
dimensional information of a subject. In a field of film appreciation, there is
appreciation media for a three-dimensional picture such as a print which is made by
using images acquired from the respective imaging units and having parallaxes, and is
implemented by using a lenticular lens sheet. The compound eye camera is used as an
image input device for the media.
As described above, the compound eye camera includes the plurality of imaging
units. Accordingly, favorable parallax images cannot be acquired under a condition in
which there are variations in characteristics of a lens, an imaging element or the like of
each imaging unit or variations in the placement attitude on an imaging unit-by-imaging
unit basis when mounting the imaging units. In particular, the latter variations greatly
influence the parallax information. Thus, it is important to reduce the variations in
placement attitudes of the imaging units for the sake of prevention of degrading a
stereoscopic effect and viewability of the three-dimensional image.
When a subject (person) SUB2 is imaged by two imaging units (left and right)
as shown in Fig. 12A, it is difficult to acquire ideal parallax images. The ideal parallax

images IMGL and IMGR are imaged under a condition without variations in the attitude
of the imaging units in the camera, and between the ideal parallax images IMGL and
IMGR, the image SUB2L of the subjects SUB in the image IMGL imaged by the left
imaging unit and the image SUB2R of the subject SUB2 in the image IMGR imaged by
the right imaging unit are displaced each other toward a horizontal direction of the
images IMGL and IMGR, the displacement being due to the parallax, as shown in Fig.
12B. In actually-acquired parallax images IMGL' and IMGR', the image SUB2L' and
SUB2R' of the subject SUB2 in the images IMGL' and IMGR' are displaced due not only
to parallax, but also to the attitudes of the imaging units (Fig. 12C).
In a process of manufacturing me compound eye camera, it is preferable that the
imaging units are mounted with high location precision so as to eliminate the above-
mentioned displacement of the image of the subject due to the variations of the attitudes
of the imaging units. However, the pursuit of precision causes problems of reduction in
productivity, increase in operation worker-hour and the like.
In order to solve the above described problems, conventionally, PTL 1 and PTL
2 suggest methods which eliminate a difference in geometrical characteristics between
the imaging units by geometrically transforming images taken by the compound eye
camera.
In a method described in PTL 1, in order to correct projective distortions caused
by displacement in positions and attitudes of the plurality of cameras, images taken by
the respective cameras are projected on the same (single, identical) plane, and the images
taken by the respective cameras are corrected to be the images on the same plane.
A method described in PTL 2 is suggested to correct deviations caused by
differences in attitudes and zoom factors of the imaging units among deviations in angles
of view which arise when two imaging units simultaneously take images of a subject
In the method described in PTL 2, corresponding points between images acquired by the
respective imaging units are detected, calculates a geometrical correction parameter such
that the corresponding points of respective images are identical to each other, and
transforms the images using the parameter.
Citation List
Patent Literature
PTL 1: Japanese Patent Application Laid-Open No. 7-294215

PTL 2: Japanese Patent Application Laid-Open No. 2006-157432
Summary of Invention
Technical Problem
However, in the method described in PTL 1, image processing that forcedly
projects the individual images on the identical plane is performed. Therefore, the
method described in PTL 1 has a problem that the parallax between the images after
projection becomes inappropriate and out of intention.
On the other hand, since in the method described in PTL 2, the images are
transformed such that the corresponding points in the two images are identical to each
other. The method described in PTL2 has a problem that the displacement of the image
of the subject between the parallax images due to a parallax to be left is also corrected,
and information on the parallax is lost.
The presentiy disclosed subject matter is made in view of these situations. It is
an object of the presently disclosed subject matter to provide a stereoscopic image
processing device, a stereoscopic image processing method, a recording medium and a
stereoscopic imaging apparatus which, when eliminating variations in views of angle
between a plurality of images due to variations in placement attitudes, zoom factors and
the like of a plurality of imaging units which take the plurality of images by a
geometrical transformation, can geometrically transform the images so as to leave the ,
intrinsical deviation of view of angle to between the images, can acquire ideal parallax
images.
Solution to Problem
In order to achieve the above-mentioned object, a stereoscopic image processing _
device concerning a first aspect of the presently disclosed subject matter, includes an
image acquisition device for acquiring a plurality of images of an identical subject taken
from a plurality of viewpoints; a corresponding point detection device for selecting a
prescribed image as a reference image from among the acquired plurality of images,
selecting an image other than the reference image as a target image from among the
acquired plurality of images, and detecting a plurality of feature points from the
reference image and a plurality of corresponding points from the target image to generate

a plurality of pairs of the feature point and corresponding point, wherein feature of the
feature point and the corresponding point included in the same pair are substantially
identical to each other; a parameter estimation device for estimating geometrical
transformation parameters for geometrically transforming the target image such that y
coordinate values of the feature point and the corresponding point included in the same
pair are substantially identical to each other, wherein y direction is orthogonal to a
parallax direction of the plurality of viewpoints; and an image transformation device for
geometrically transforming the target image on the basis of the estimated geometrical
transformation parameters.
In the first aspect of the presently disclosed subject matter, the geometrical
transformation parameters for geometrically transforming the target image are estimated
under a constraint that a component along a direction orthogonal to a parallax direction
of the plurality of viewpoints be substantially zero with respect to each displacement
between the feature points of a prescribed image (reference image) among the plurality
of images of the identical subject taken from the plurality of viewpoints and the
corresponding points of the target image other than the reference image. More
specifically, the apparatus of the first aspect estimates the geometrical transformation
parameter for geometrically transforming the target image such that only the y coordinate
values of the feature point and the transformed corresponding points are substantially
identical to each other. Geometrical transformation based on the thus estimated
geometrical transformation parameter can correct a displacement in an angle of view
along the direction orthogonal to the parallax direction, while remaining an intrinsically
displacement in the angle of view to between the reference image and the target image,
thereby allowing ideal parallax images to be acquired.
A second aspect of the presently disclosed subject matter provides a
stereoscopic image processing device according to the first aspect, wherein the parameter
estimation device estimates at least some parameters in the geometrical transformation
parameters based on at least the x and y coordinate values of the corresponding points
and the y coordinate values of the feature points.
A third aspect of the presently disclosed subject matter provides a stereoscopic
image processing device according to the second aspect, wherein the parameter

estimation device calculates parameters other than said some parameters estimated by the
parameter estimation device based on said some parameters.
More specifically, there are a case where all the parameters can be estimated
using the x and y coordinate values of the corresponding points and the y coordinate
values of the feature points, and a case where some of the parameters, not all the
parameters can be estimated, depending on the types of transformation equations for
performing the geometrical transformation. However, even in the latter case, the
parameter other than said some parameters can be calculated based on said some
parameters.
A fourth aspect of the presently disclosed subject matter provides a stereoscopic
image processing device according to the second aspect, wherein the geometrical
transformation parameters are a projective transformation parameters, and the parameter
estimation device selects five or more pairs from among the plurality of pairs of the
feature point and the corresponding point, and estimates parameters for deterrnining y
coordinate values of projective-transformed corresponding points on the basis of the
coordinate values the feature point and the corresponding point of each of the selected
pairs.
A fifth aspect of the presently disclosed subject matter provides a stereoscopic
image processing device according to the fourth aspect, wherein the parameter estimation
device calculates the other parameters required to determine the x coordinate values of
the projective-transformed corresponding points on the basis of the parameters for
determining y coordinate values of projective-transformed corresponding points.
More specifically, when the geometrical transformation is performed by the
projective transformation, only some of the parameters (parameters for at least
determining the y coordinate values of the projective-transformed corresponding points)
can be estimated among the projective transformation parameters on the basis of the x
and y coordinate values of the corresponding points the y coordinate values of the feature
points with respect to five or more pairs. However, parameters other than said some
parameters (parameters required to determine the x value of the projective-transformed
corresponding points) can be calculated on the basis of the estimated parameters.
A sixth aspect of the presently disclosed subject matter provides a stereoscopic
image processing device according to the second aspect, wherein the geometrical

transformation parameters are Helmert transformation parameters, and the parameter
estimation device selects three or more pairs from among the detected plurality of pairs
of the feature point and the corresponding point, and estimates the Helmert
transformation parameters on the basis of the coordinate values of the feature points and
the corresponding points of each of the selected pairs.
More specifically, when the geometrical transformation on the image is
performed by the Helmert transformation, all of Helmert transformation parameters can
be estimated on the basis of the x and y coordinate values of the corresponding points
and the y coordinate values of the feature points with respect to at least three pairs.
A stereoscopic image processing method concerning a seventh aspect of the
presently disclosed subject matter includes an image acquisition step of acquiring a
plurality of images of an identical subject taken from a plurality of viewpoints; a
corresponding point detection step of selecting a prescribed image as a reference image
from among the acquired plurality of images, selecting an image other than the reference
image as a target image from among the acquired plurality of images, and detecting a
plurality of feature points from the reference image and a plurality of corresponding
points from the target image to generate a plurality of pairs of the feature point and
corresponding point, wherein feature of the feature point and the corresponding point
included in the same pair are substantially identical to each other; a parameter estimation
step of estimating geometrical transformation parameters for geometrically transforming
the target image such that y coordinate values of the feature points and the corresponding
points included in the same pair are substantially identical to each other, wherein y
direction is orthogonal to a parallax direction of the plurality of viewpoints; and an image
transformation step of geometrically transforrning the target image on the basis of the
estimated geometrical transformation parameters.
A eighth aspect of the presently disclosed subject matter provides a stereoscopic
image processing method according to the seventh aspect, wherein, in the parameter
estimation step, at least some parameters in the geometrical transformation parameters
are estimated based on at least the x and y coordinate values of the corresponding points
and the y coordinate values of the feature points.
A ninth aspect of the presently disclosed subject matter provides a stereoscopic
image processing method according to the eighth aspect, wherein the parameter

estimation step includes: a first step of randomly selecting a certain number of pairs
required to estimate the geometrical transformation parameters for determining y
coordinate values of the transformed corresponding points from among N pairs, provided
that a total number of plurality of pairs is N; a second step of calculating the y coordinate
values of the transformed corresponding points with respect to each of the N pairs based
on the parameter estimated on the basis of the coordinate values of each corresponding
point randomly selected by the first step; a third step of calculating a difference between
the y coordinate value of the transformed corresponding point calculated by the second
step and the y coordinate value of the feature point for each of the N pairs; a fourth step
of counting a number of pairs of the feature point and the corresponding point whose
difference calculated by the third step is less than a predetermined first threshold; a fifth
step of determining a confidence level of the estimated parameter on the basis of a ratio
between the counted number of pairs and the N; and a step of iterating the first to fifth
steps until the determined confidence level reaches a predetermined confidence level, or
the number of iterations reaches a predetermined number of iterations.
According to the ninth aspect of the presently disclosed subject matter, an
appropriate combination can be determined from among the N pairs as the certain
number of pairs (five pairs for the projective transformation, and three pairs for the
Helmert transformation) required to estimate the parameters for at least determining the y
coordinate values of the transformed corresponding points. More specifically, the
certain number of pairs required to estimate the parameter from among the N pairs are
randomly selected. The y coordinate value of the transformed corresponding point is
calculated with respect to each of the N pairs based on the estimated parameter on the
basis of the coordinate value of each corresponding point of the selected pairs. The
difference between the calculated y coordinate value of the transformed corresponding
point and the y coordinate value of the feature point is calculated with respect to each of
the N pairs. The number of pairs of the feature point and the corresponding point
whose calculated difference is less than the first threshold is counted. The confidence
level of the estimated parameter is determined on the basis of the ratio of the number of
counted pairs and the N. The processes including the random selection of pairs and the
like are iterated until the determined result of the estimated parameter reaches the
predetermined confidence level or the number of iterations reaches the predetermined

number of iterations. Thus, the appropriate combination can be determined as the
certain number of pairs required to estimate the parameter for at least determining the y
coordinate value after transformation.
A tenth aspect of the presently disclosed subject matter provides a stereoscopic
image processing method according to the ninth aspect, wherein the parameter estimation
step includes: a sixth step of calculating the y coordinate value of the transformed
corresponding point for each of the N pairs based on the estimated parameter when the
determined confidence level reaches the predetermined confidence level or the estimated
parameter when the confidence level is highest among the levels at the respective
iterations; a seventh step of calculating a difference between the y coordinate value of the
transformed corresponding point calculated by the sixth step and the y coordinate value
of the feature point for each of the N pairs; an eighth step of selecting only the pairs of
the feature point and the corresponding point whose difference calculated by the seventh
step is less than a predetermined second threshold from the N pairs; and a ninth step of
calculating the plurality of parameters using only the pairs of the feature point and the
corresponding point selected in the eighth step.
More specifically, the estimated parameters have a desired confidence level.
The y coordinate value of the transformed corresponding point transformed by this
parameters and the y coordinate value of the feature point should intrinsically be
substantially identical, and the difference thereof is less than the predetermine second
threshold. On the other hand, the pair of the feature point and the corresponding point
whose difference is less than the second threshold regarded as inappropriate pair and
eliminated. Only the pair of corresponding points whose difference is less than the
second threshold is extracted. The plurality of parameters are then calculated again
based on the pair thus extracted, thereby allowing the confidence level of the parameter
to further increase.
A eleventh aspect of the presently disclosed subject matter provides a
stereoscopic image processing method according to the tenth aspect, wherein, in the
ninth step, the plurality of parameters which minimize a square sum of differences
between the y coordinate values of the transformed corresponding points in the plurality
of pairs selected by the eighth step and the y coordinate values of the feature points. In
other words, the plurality of final parameter is estimated by the least squares method.

A recording medium concerning a twelfth aspect of the presently disclosed
subject matter includes a computer program causing a computer to execute a process for
a stereoscopic image processing, the process including: acquiring a plurality of images of
an identical subject taken from a plurality of viewpoints; selecting a prescribed image as
a reference image from among the acquired plurality of images, selecting an image other
than the reference image as a target image from among the acquired plurality of images,
and detecting a plurality of feature points from the reference image and a plurality of
corresponding points from the target image to generate a plurality of pairs of the feature
points and corresponding points, wherein features of the feature points and the
corresponding points included in the same pair are substantially identical to the feature
points respectively; estimating geometrical transformation parameters for geometrically
transforming the target image such that y coordinate values of the feature points and the
corresponding points included in the same pair are substantially identical to each other,
wherein y direction is orthogonal to a parallax direction of the plurality of viewpoints;
and geometrically transforming the target image on the basis of the estimated
geometrical transformation parameters.
A stereoscopic imaging apparatus concerning a thirteenth aspect of the presently
disclosed subject matter includes a stereoscopic image processing device according to
any one of the first to the fifth aspect; and a plurality of imaging units which are disposed
at a plurality of viewpoints along the parallax direction respectively, and take images of
the identical subject from their viewpoints, wherein the image acquisition device
acquires the plurality of images taken by the plurality of imaging units, respectively.
The above described aspects of the presently disclosed subject matter may be
provided as a stereoscopic image processing program which causes an apparatus such as
a computer or a processing unit (CPU) in a camera, an image reproduction apparatus or a
printer to execute the above described procedures. Also, the above aspects of the
presently disclosed subject matter may be provided as a computer program product in a
computer-readable recording medium for use in controlling the apparatus.
Advantageous Effects of Invention
The above described aspects of the presently disclosed subject matter estimates
the geometrical transformation parameter for geometrically transforming a target image

under a constraint making the component along the direction orthogonal to the parallax
direction which depends on the positional relationship between the plurality of
viewpoints be about substantially zero with respect to the displacements between the
feature points on the prescribed image (reference image) among the plurality of images
of the identical subject taken from the plurality of viewpoints and the respective
corresponding points on the image (the target image) other than the reference image
among the plurality of images. Accordingly, the aspects of the presently disclosed
subject matter can correct the deviation in angle of view along the direction orthogonal to
the parallax direction, while leaving the deviation in angle of view which intrinsically
arises along the parallax direction. Thus, the ideal parallax images can be acquired.
Brief Description of the Drawings
Fig. 1 is a block diagram illustrating a construction of an embodiment of a
stereoscopic imaging apparatus;
Fig. 2 is a block diagram illustrating a construction of an imaging unit shown in
Fig. 1;
Fig. 3 is a diagram illustrating a positional relationship between a disposition of
a plurality of imaging units (six imaging units) in an apparatus body of the stereoscopic
imaging apparatus and a subject;
Fig. 4A is a diagram illustrating an example of ideal parallax images taken by
the six imaging units;
Fig. 4B is a diagram illustrating an example of actual parallax images taken by
the six imaging units;
Fig. 5 is a flowchart illustrating a flow of a stereoscopic image processing
according to an embodiment of the presently disclosed subject matter,
Fig. 6A is a diagram illustrating a correspondence relationship between feature
points in a reference image and corresponding points in a target image;
Fig. 6B is a diagram illustrating a vector connecting a feature point and a
corresponding point corresponding to the feature point;
Fig. 7A is a diagram illustrating a vector connecting the feature point and the
corresponding point and vector components thereof;

Fig. 7B is a diagram illustrating an embodiment of estimating method of
estimating a projective transformation parameter;
Fig. 7C is a diagram illustrating a comparative example;
Fig. 8 is a flowchart illustrating an embodiment of a procedure for estimating a
projective transformation parameter (No. 1);
Fig. 9 is a flowchart illustrating the embodiment of the procedure for estimating
the projective transformation parameter (No. 2);
Fig. 10 is a flowchart illustrating the embodiment of the procedure for
estimating the projective transformation parameter (No. 3);
Fig. 11 is a diagram illustrating a procedure for determining other parameters
required for determining an x coordinate value after the projective transformation among
the projective transformation parameters;
Fig. 12A is a diagram illustrating an example of a disposition of two imaging
units (left and right) and a subject;
Fig. 12B is a diagram illustrating an example of ideal parallax images; and
Fig. 12C is a diagram illustrating a deviation between the images due to camera
attitudes of the imaging units.
Description of Embodiments
A stereoscopic image processing device, a stereoscopic image processing
method, a recording medium and a stereoscopic imaging apparatus according to
embodiments of the presently disclosed subject matter will hereinafter be described with
reference to the accompanying drawings.
[Overall Configuration of Stereoscopic Imaging Apparatus]
Fig. 1 is a block diagram illustrating a construction of an embodiment of a
stereoscopic imaging apparatus of the presently disclosed subject matter.
As shown in Fig. 1, the stereoscopic imaging apparatus 10 includes six imaging
units 1 to 6. The stereoscopic imaging apparatus 10 acquires six images (parallax
images) by taking the identical subject from six viewpoints, and records the images as
image data for recording in a prescribed format.
To a central processing unit (CPU 12), the imaging units 1 to 6 and a light
emitter 18 are connected via a control bus 16. A main memory 20, a digital signal

processing section 22, an integrating accumulator 24, a compression/expansion
processing section 26, an external recording section 28, a display section 30, a
corresponding point detecting section 32, a geometrical transformation section 34, and a
geometrical transformation parameter estimation section 36 are also connected to the
CPU 12. The CPU 12 controls the operation of the stereoscopic imaging apparatus 10
on the basis of an input operation from the operation section 14 according to a prescribed
control program.
The imaging units 1 to 6, the main memory 20, the digital signal processing
section 22, the integrating accumulator 24, the compression/expansion processing section
26, the external recording section 28, the display section 30, the corresponding point
detecting section 32, the geometrical transformation section 34, and the geometrical
transformation parameter estimation section 36 are connected to each other via a data bus
38.
The constructions of the imaging units 1 to 6 are the same as each other. As
shown in Fig. 2, each of the imaging units 1 to 6 includes an imaging lens 40, a
diaphragm 41, an IR (infrared) cut filter 42, an optical low-pass filter 43, an imaging
element (CCD (charge coupled device) 44), an A/D converter 45, a lens driver 46, a
diaphragm driver 47, and a CCD driver 48.
The imaging lens 40 includes a focusing lens, a zoom lens, is driven by the lens
driver 46 and moves back and forth along an optical axis thereof. The CPU 12 controls
the position of the focusing lens to adjust the focus to be on a subject by controlling the
lens driver 46. The CPU 12 controls a zooming by controlling the position of the zoom
lens according to a zoom instruction from the operation section 14.
The diaphragm 41 includes, for instance, an iris diaphragm. The diaphragm 41
is driven by the diaphragm driver 47. The CPU 12 controls the amount of aperture
(aperture value) of the diaphragm 41 via the diaphragm driver 47, and controls the
amount of incident light into the CCD 44.
The CCD 44 is a two-dimensional color CCD solid-state imaging element. The
CCD 44 includes multiple photodiodes which are two-dimensionally arranged on a
photo-receptive surface of the CCD 44, and color filters (for example, R (Red), G
(Green) and B (Blue) filters) which are disposed on the respective photodiodes in a
prescribed arrangement. An optical image formed on the photo-receptive surface of the

CCD 44 via the imaging lens 40, the diaphragm 41, the IR cut filter 42 and the optical
low-pass filter 43 is converted into signal charges corresponding to the amount of
incident light by these photodiodes. The signal charges accumulated in the respective
photodiodes are successively read out from the CCD 44 as voltage signals (R, G and B
image signals) corresponding to the amount of the signal charges on the basis of drive
pulses provided from the CCD driver 48 according to an instruction by the CPU 12.
The CCD 44 is provided with an electronic shutter function. The CPU 12 controls a
charge accumulating time during which a charge is accumulated into the photodiodes by
the electronic shutter function. That is, the CPU 12 controls an exposure time or a
shutter speed by the electronic shutter function. Although the CCD 44 is used as the
imaging element in this embodiment, an imaging element having another configuration
such as a CMOS (complementary metal-oxide semiconductor) sensor and the like can be
used.
The image signal read out from the CCD 44 is converted into a digital signal by
the A/D converter 45. Subsequently, the digital signal (image data) is temporarily
stored in the main memory 20 via the data bus 38.
As shown in Fig. 3, the imaging units 1 to 6 are disposed along a horizontal
direction to a main body 11 of the stereoscopic imaging apparatus 10 at prescribed
intervals (prescribed base line lengths). The imaging units 1 to 6 are disposed so as to
adjust convergent angles (angles between optical axes of the imaging units) of the
imaging units 1 to 6 such that the optical axes of the imaging lenses of the respective
imaging units 1 to 6 intersect at a single point.
The CPU 12 drives the imaging units 1 to 6 in synchronism with each other.
More specifically, the CPU 12 always adjusts focus of each imaging lens 40 included In
imaging units 1 to 6 so that the identical subject comes into focus. The imaging units 1
to 6 are always set to have the same focal length (zoom factor). Furthermore, the CPU
12 adjusts the diaphragm 41 so that each CCD 44 included in the imaging units 1 to 6
always obtain the same amount of incident light (aperture value).
The operation section 14 includes devices for receiving the user input such as a
shutter button, a power switch, a mode dial, a cross button, a zoom button. The shutter
button is pressed in two stages, which is put into so-called "half-pressing state" in which
the shutter button is pressed halfway (half-pressed) and so-called "full-press state" in

which the shutter button is fully pressed. In an image-taking mode, when the shutter
button is pressed halfway, an image-taking preparation process (for example, Automatic
Exposure process (AE), Automatic Focus adjustment process (AF) and/or Automatic
White Balance correction process (AWB) is performed. In an image-taking mode,
when the shutter button is fully-pressed, an image-taking and recording process is
performed.
The light emitter 18 includes, for instance, a discharge tube (xenon tube). The
light emitter 18 emits light, if necessary, for instance, when taking an image of a dark
subject, or taking an image in a back light situation.
The main memory 20 is used as a working area when the CPU 12 executes a
program, and as a storage section for temporarily storing digital image signals acquired
by imaging by the imaging units 1 to 6.
The digital signal processing section 22 includes a white balance adjustment
circuit, a gradation transformation circuit (e.g. gamma correction circuit), a color-
interpolating circuit (a processing circuit for obtaining color signals corresponding to
colors other than a color of the color filter disposed on a pixel by interpolating the color
signals obtained from the neighboring pixels, and obtaining the RGB color signals for
each pixel position.), a contour correction circuit, a brightness/color-difference signal
generating circuit. The digital signal processing section 22 performs a prescribed signal
processing on the R, G and B image data stored in the main memory 20. More
specifically, the R, G and B image signals are converted by the digital signal processing
section 22 into YUV signals including brightness signals (Y signals) and color-difference
signals (Cr and Cb signals), and subjected to prescribed processes such as the gradation
transformation process (e.g. gamma correction). The image signal processed by the
digital signal processing section 22 is stored in the main memory 20.
The integrating accumulator 24 calculates a focus evaluation value used for the
automatic focus adjustment process on the basis of an image signal taken when the
shutter button is pressed halfway. Also, the integrating accumulator 24 calculates
brightness of the subject required for the automatic exposure process. In the automatic
focus adjustment process, the CPU 12 searches for a position where the focus evaluation
value calculated by the integrating accumulator 24 becomes a local-maximum. The
CPU 12 makes the focusing lens move to that position, and focuses on an image of the

subject (a main subject). In the automatic exposure process, the CPU 12 performs
exposure setting for obtaining an appropriate exposure amount on the basis of the
brightness of the subject calculated by the integrating accumulator 24. More
specifically, the CPU 12 sets image-taking sensitivity, the aperture value and the shutter
speed, and judges a necessity of firing a flush (the light emitter 18).
The compression/expansion processing section 26 performs a compression
process on the inputted image data and generates compressed image data in a prescribed
format according to an instruction from the CPU 12. For instance, a still image is
subjected to a compression process conforming to the JPEG standard, and a moving
image is subjected to a compression process conforming to the MPEG2, MPEG4 or
H.264 standard. The compression/expansion processing section 26 also performs a
expansion process on inputted compressed image data and generates uncompressed
image data according to an instruction from the CPU 12.
The external recording section 28 records the image file containing the image
data in the JPEG format or the like generated by the compression/expansion processing
section 26 into a detachable external recording medium such as a memory card. Also,
the external recording section 28 reads the image file from the external recording
medium.
The display section 30 includes, for instance, a color liquid crystal panel. The
display section 30 displays an image having been taken by the stereoscopic imaging
apparatus 10, and is used as a GUI (Graphical User Interface) for various settings. The
display section 30 is also used as an electronic view finder for a user to confirm the angle
of view in an image-taking mode. The display section 30 includes a lenticular lens
including a group of semi-cylindrical lenses which is disposed on a surface of the color
liquid crystal panel. In a playback mode for playing back a stereoscopic image (3D
images) on the basis of images from a plurality of standpoints (parallax images), the
display section 30 displays the parallax images read out from the external recording
section 28, and allows a user to stereoscopically view the images. Meanwhile,
examples of a device for playing back a stereoscopic image on the display section 30
include, but are not limited to the lenticular lens. For example, so-called parallax
barrier system can be applied to the display section 30. The parallax barrier system
controls the display section 30 to alternately repeat a process for displaying an image for

a left eye of a user, which is made by the parallax images, on the color liquid crystal
panel and emitting the panel by a backlight so that the light emitted by the backlight
reaches only for the left eye of the user by using a parallax barrier, and a process for
displaying an image for a right eye of the user, which is made by the parallax images, on
the panel and emitting the panel by the backlight so that the light emitted by the
backlight reaches only for the right eye of the user by using the parallax barrier.
As shown in Fig. 4A, in an ideal case, the images SUB1-1 to SUB1-6 of the
subject SUB1 only by the parallaxes between six images acquired by the imaging units 1
to 6 respectively. However, as shown in Fig. 4B, in an actual case, the images SUB1-1'
to SUB 1-6' of the subject SUB1 are displaced each other due to variations in attitudes
and zoom factors of the imaging units 1 to 6 in addition to the parallaxes.
The images are transformed to correct the above-mentioned displacement
concerning the subject by using the corresponding point detecting section 32, the
geometrical transformation section 34 and the geometrical transformation parameter
estimation section 36. A description related to a process for correcting the above-
mentioned deviations will hereinafter be made with reference to a flowchart shown in Fig.
5.

Fig. 5 is a flowchart illustrating a flow of a stereoscopic image processing
according to an embodiment of the presently disclosed subject matter.
[StepS 10]
When taking parallax images, six imaging units 1 to 6 take images on the
identical subject, and the acquired six images are recorded. Here, each image is
subjected to the above-mentioned various types of signal processing by the digital signal
processing section 22, and subsequendy, is temporarily stored in the main memory 20.
Instead, it is also preferable that the external recording section 28 records the image into
the external recording medium.
[StepS 12]
A prescribed one image (image acquired by the imaging unit 3 in this
embodiment) among the six images stored in the main memory 20 is set as a reference
image. The set reference image is read from the main memory 20. Although the
image acquired by the imaging unit 3 is set as the reference image in this embodiment, it

does not mean any limitation. Instead, any image among the six images can be set as
the reference image.
[StepS 14]
The prescribed one image is selected among the five images other than the
reference image as a target image. The selected target image is read from the main
memory 20.
[Step S16]
The corresponding point detecting section 32 detects a plurality of pairs of
corresponding points whose features are substantially identical to each other with respect
to the above-mentioned set reference image and the selected target images.
Conventionally, various methods have been proposed as the detection method of
the corresponding points. Conventional arts such as a block matching method, a KLT
method (Tomasi & Kanade, 1991, Detection and Tracking of Point Features), SIFT
(Scale Invariant Feature Transform) and the like can be used as the detection method.
In this embodiment, for convenience sake, the corresponding point detected
from the reference image among the pairs of corresponding points detected by the
corresponding point detecting section 32 is referred to as a feature point.
[Step S18]
The geometrical transformation parameter estimation section 36 estimates and
calculates a projective transformation parameter on the basis of the coordinate values of
the plurality pairs of the feature point and the corresponding point. The details on the
step S18 will hereinafter be described.
[Step S20]
The CPU 12 determines whether or not the projective transformation parameter
is estimated successfully. When the parameter is estimated successfully (in a case of
"yes"), transition to step S22 is made. When the parameter is estimated unsuccessfully
(in a case of "no"), step S22 is skipped and transition to step S24 is made.
[Step S22]
The geometrical transformation section 34 projective-transforms the target
image on the basis of the projective transformation parameter having been estimated
successfully. The projective-transformed image is recorded by the external recording
section 28 into the external recording medium.

[Step S24]
The CPU 12 determines whether or not the processing of the above-mentioned
steps S14 to S22 between the above-mentioned reference image and each of the five
target images. When the processing is not finished, transition to step S14 is made and
selection of another target image is performed in the step. When the processing is
finished, this stereoscopic image processing is finished.

Next, a procedure for estimating the projective transformation parameter on the
basis of the plurality pairs of the feature point and the corresponding point will be
described.
Here, projective transformation equations are as follows:
[Expression 1]
X = (ax + by + s)/(px + qy + 1)
Y = (cx + dy + t)/(px + qy + 1)
The projective transformation parameters are represented as eight parameters: a,
b, s, c, d, t, p and q in [Expression 1]. (x, y) and (X, Y) represent coordinate values
before and after the projective transformation, respectively.
In Fig. 6A, the feature points (solid dots) extracted from the reference image
IMG1 are superimposed onto the reference image IMG1. Also, the corresponding
points (open dots) detected from the target image IMG2 are superimposed onto the target
image IMG2.
Fig. 6B illustrates vectors connecting the feature points and the respective
corresponding points. In Fig. 6B, the feature point A (xl, yl) and the corresponding
point A' (x2, y2) represent a certain pair of the feature point and the corresponding point.
Intrinsically, the feature point A and the corresponding point A' should have the
same height (i.e. the y coordinate values are substantially the same as each other).
However, since the subject in the reference image IMG1 and the target image IMG2 is
displaced due to variations in attitudes and zoom factors of the imaging units 1 to 6, the
heights (y coordinate values) of the feature point A and the corresponding point A' are
not identical to each other.
Thus, when estimating the projective transformation parameter, this embodiment
provides with a constraint that a line segment (hereinafter referred to as "vector")

connecting the feature point A and the corresponding point A' becomes substantially
horizontal (i.e. y coordinate values of the feature point A and the corresponding point A'
are substantially identical to each other). This constraint is based on a fact that the six
imaging units 1 to 6 are disposed horizontally to the main body 11 of the apparatus 10.
If a plurality of imaging units are vertically disposed, an assumption that the vectors are
vertical is required (i.e. x coordinate values of the feature point and the corresponding
point are substantially identical to each other). . In other words, a constraint that
coordinate values along a direction orthogonal to parallaxes of the plurality of viewpoints
are substantially identical with each other is provided.
Fig. 7A is a diagram illustrating a vector connecting the feature point and the
corresponding point and vector components thereof. Fig. 7B is a diagram illustrating an
embodiment of estimating method of estimating a projective transformation parameter.
Fig. 7C is a diagram illustrating a comparative example.
As shown in Fig. 7A, the vector (combined vector Vc in Fig. 7A) connecting the
feature point and the corresponding point can be separated into a plurality of vector
components. The vector components represent some factor which causes a
displacement between the feature point and the corresponding point, and include
"parallax" and "factors other than the parallax such as a variation in attitude of the
imaging unit and the like". In Fig. 7A, a vector Vp represents a vector component
corresponding to the parallax, and a vector Vj represents a vector component
corresponding to the factors other than parallax including the variation in attitude of the
imaging unit 1 to 6. As understood from the Fig. 7A, the y component of the combined
vector Vc does not depend on the vector component corresponding to the parallax, and
the y component of the combined vector Vc only depend on the factors other than the
parallax.
As shown in Fig. 7C, when the target image IMG2 is projective-transformed so
that the feature point and the corresponding point are substantially identical with each
other, by a correction vector VA' (~ -Vc), the combined vector in the projective-
transformed image becomes zero.
Thus, the presently disclosed subject matter estimates and calculates the ideal
projective transformation parameter related to the factors other than the parallax
including the variation in camera attitude of the imaging unit 1 to 6 based on the y

component of the vector Vc. As shown in Fig. 7B, when the target image IMG2 is
projective-transformed so as to eliminate the vector component related to the factors
other than the parallax by a correction vector VA (~ -V1), the combined vector in the
projective-transformed image only include the vector component VP related to the
parallax.

Figs. 8 to 10 are flowcharts illustrating an embodiment of a procedure for
estimating a projective transformation parameter
[Step S100]
Provided that the number of all pairs of the feature points extracted from the
reference image IMG1 and the corresponding points which are detected in the target
image EMG2, and correspond to the respective feature points is N, the N pairs of the
coordinates (x, y) of the feature points and the coordinates (X, Y) of the corresponding
points are read. It is also provided that the coordinate of the feature point in an i-th (1 <
i < N) pair is (xi, yi) and the coordinate of corresponding point thereof is (Xi, Yi).
[StepS 102]
Since the projective transformation parameter cannot be estimated when the
number of pairs N of the feature points and the corresponding points acquired by the
corresponding point detecting section 32 is small, a threshold process on the number of
pairs N is performed.
More specifically, a threshold value TH1 on the number of pairs N of the feature
points and the corresponding points is provided, and following determination process is
performed.
IF N < TH1 then A flag indicating that the estimation of the parameter fails is
set, and the estimation is finished (step S124 in Fig. 10).
ELSE then The estimation of the parameter is continued.
In the embodiment, at least five pairs of the coordinate values are required in
order to estimate the projective transformation parameter. Therefore, the threshold
value TH1 is a prescribed value of at least 5.
[StepS 104]
The number of iterations "irand" and a maximum value "n_vote_max" of
confidence level parameter are initialized (both of them are set to "0").

[StepS 106]
Five pairs are randomly selected from among the N pairs of the feature point and
the corresponding point. The random selection of the five pairs can be performed using
random numbers.
[StepS 108]
The projective transformation parameter for matching the y coordinates of the
five pairs of the feature point and the corresponding point with each other is calculated
on the basis of the coordinate values (the x and y coordinate values of the five feature
points and y coordinate value of the five corresponding points) of the five pairs of the
feature point and the corresponding point selected in step S106.
As shown in the above-mentioned equation [Expression 1], the projective
transformation parameters with reference to the y direction are the five, or c, d, t, p and q.
Accordingly, the parameters (c, d, t, p and q) can be calculated uniquely by solving
simultaneous equations which are acquired by substituting the coordinate values of the
five points for the equation [Expression 1].
[Step S110]
The all pair (N pair) of points are applied to the projective transformation
equation [Expression 1] for which the above-mentioned calculated parameters (c, d, t, p
and q) have been substituted, and the number "n_vote" of pair of points which satisfy a
following conditional expression is counted.
[Expression 2]
Conditional Expression 1: |Yi - yi'| < THY1
where yi' = (cx; + dyi + t)/(pXi + qyi + 1), THY1 is a prescribed constant (threshold).
The conditional expression 1 is used to determine whether or not the y
coordinate values of another pair of the points are matched, on the basis of the projective
transformation parameter determined from the five pairs of the points. More
specifically, the value of "n_vote" represents how many vectors are horizontal among the
N vectors. The larger the "n_vote" is, the higher the confidence level of the projective
transformation parameter is. The parameter "n_vote" is referred to as a confidence
level parameter hereinafter.
[Step S112]

The confidence level parameter "n_vote" calculated in step S110 and the
maximum value "n_vote_max" of the confidence level parameter are compared
(Determination Process of Confidence Level 1). When the confidence level parameter
n_vote is larger than n_vote_max (n_vote > n_vote_max), transition to step S114 is made.
When n_vote is equal to or smaller than n_vote_max (n_vote < n_vote_max), transition
to step S118 is made.
[Step S114]
The calculated parameters (c, d, t, p and q) are temporarily stored as an
intermediate parameters (c_tmp, d_tmp, t_tmp, p_tmp and q_tmp), and n_vote is
temporarily stored as the maximum value n_vote_max.
When the projective parameters are estimated, the processing of steps S106 to
S118 are iterated prescribed times "nrand" as will be described later. The present
embodiment has an assumption that the parameters (c, d, t, p and q) corresponding to the
largest n_vote has the highest confidence level. Thus, the confidence level parameter
n_vote which is calculated by a previous calculation of steps S106 to S120 (the (irand -
l)-th calculation) is stored as n_vote_max. When the confidence level parameter
n_vote calculated by a last calculation is larger than the confidence level parameters
calculated by the previous calculation, which is stored as n_vote_max, the above
mentioned parameters (c_tmp, d_tmp, t_tmp, p_tmp, q_tmp and n_vote_max) are
updated.
[Step S116]
A ratio of the maximum value n_vote_max of the confidence level parameter to
the number N of all pairs (n_vote_max/N) is acquired, and a threshold process on the
ratio is performed (Determination Process of Confidence Level 2).
More specifically, a following process is performed on the basis of a prescribed
threshold TH2.
IF n_vote_max/N < TH2 then The process for estimating the projective
transformation parameters is continued.
ELSE then The iteration of the process for estimating the parameter is finished,
and transition is made for calculating a final projective transformation parameter (step
S126 in Fig. 10).

The above-mentioned threshold TH2 can be set as a value close to " 1" but
smaller than "1".
The maximum value n_vote_max of the confidence level parameter is the
number of pairs of points whose y coordinate values are substantially identical to each
other by the projective transformation performed by using the estimated projective
transformation parameters among the N pairs of points. Accordingly, n_vote_max/N
represents a ratio of the number of points to be appropriately corrected to the number of
pairs of the all points N. Therefore, when n_vote_max/N is equal to or larger than TH2,
the estimated projective transformation parameters are sufficiently confident, and the
iteration of the process for estimating the parameters is finished. Transition to
calculation of the final project transformation parameters is made. On the other hand,
when n_vote_max/N is less than TH2, the iteration of the process for estimating the
parameter is continued.
[Step S118]
It is determined whether or not the number of iterations irand is larger than the
prescribed number nrand. When the number of iterations irand is less than or equal to
the prescribed number nrand (in a case of "no"), transition to step S120 is made. When
the number of iterations is larger than the prescribed number nrand ("in a case of yes"),
transition to step S122 is made. The prescribed number nrand depends on the number
of pairs N of the feature points and the corresponding points in the image. The
prescribed number nrand is about 1000 to 10000.
[Step S120]
The number of iterations irand is incremented by one, and transition to step
S106 is made.
[Step S122]
When the number of iterations irand is larger than the prescribed number nrand,
a ratio of the maximum value (n_vote_max) of tie confidence level parameter having
been stored in step S114 to the number of all pairs N (n_vote_max/N) is acquired, and
the threshold process is performed on this ratio (Determination Process of Confidence
Level 3).
More specifically, the following process is performed using the prescribed
threshold TH3.

IF n_vote_max/N < TH3 then A flag that indicates which the estimation of the
projective transformation parameters fails is set, and the process of the estimation is
finished (step S124 in Fig. 10).
ELSE then The iteration of the process for estimating the parameters is finished,
and transition to a step for calculating the final projective transformation parameters
(step S126 in Fig. 10) is made.
Needless to say, the threshold TH3 is less than the threshold TH2 in step S114.
[Step S124 (Fig. 10)]
As described above, when the number of pairs N of the feature points and the
corresponding points is less than the threshold value TH1 (step S102), or when the
confidence level of the estimated parameters is low (n_vote_max/N < TH3), an
estimation failure flag indicating which estimation of the projective transformation
parameters has failed is set, and the estimation of the projective transformation parameter
is finished.
[Step S126]
The intermediate parameters (c_tmp, d_tmp, t_tmp, p_tmp and q_tmp) stored in
step S114 are substituted for the projective transformation equations. The coordinate of
all pairs of the points (N pairs) are substituted for the projective transformation equations,
and only the pairs which satisfy a following conditional expression 2 are kept (the pairs
which do not satisfy the conditional expression 2 are excluded).
[Expression 3]
Conditional Expression 2: |Yi - yi'| < THY2
where yi' = (c_tmp x xi + d_tmp x y; + t_tmp)/(p_tmp x Xi + q_tmp x yi + 1), and THY2
is a predetermined constant (threshold).
Thus, only the pairs of the feature point and the corresponding point whose
confidence levels are high can be kept.
[Step S128]
The final estimation of the parameters is performed by means of a least squares
method using only the pairs of the points satisfying the above-mentioned conditional
expression 2. Here, provided that the number of the pairs of the points satisfying the
conditional expression 2 is "n", the parameters are estimated by the least squares method
so as to minimize an evaluation function "J2" in a following expression.

[Expression 4]

The evaluation function "J2" is equivalent to an evaluation function "Jl".
More specifically, as shown in following expressions, assuming that each of five
expressions acquired by partially differentiating the evaluation function J2 with respect
to each parameter (c, d, t, p and q) is "0", the parameters (c, d, t, p and q) are calculated
by solving simultaneous equations composed of these five expressions acquires.
[Expression 5]

The final parameters (c, d, t, p and q) are calculated by solving the above
simultaneous expression.
[Step S130]
Next, the residual parameters (a, b and s) among the eight projective
transformation parameters (a, b, s, c, d, t, p and q) are determined. The residual
parameters (a, b and s) are required to determine projective-transformed x coordinate
values, other than the above-mentioned estimated parameters (c, d, t, p and q).

Here, the parameter "s" corresponds to the amount of displacement along the x
direction (direction along which the parallax arises). Since the amount of parallax is
unknown and "s" cannot be uniquely determined, then "s" is assumed to be s = 0.
On the other hand, calculation of the parameters "a" and "b", for instance, is
performed as follows.
Here, assuming that a coordinate system as shown in Fig. 11, it is provided that
an image plane PL1 is disposed at a distance of "1" from the origin "O", each point on
the image plane is projected onto the projecting plane PL2. In other words, a point P1
is moved to a point P2.
Here, it is provided that the angle between the image plane PL1 and the
projecting plane PL2 is and the projecting plane PL2 is on the x' axis. It is also
provided that the point of intersection between the projecting plane and the z axis
(direction corresponding to the depth dimension with respect to the image) is O', which is
determined to be the origin on the x' axis.
Provided mat the x coordinate value of P1 is "x" and the x coordinate value of
P2 with respect to x' axis is x', a following expression is acquired.
[Expression 6]

In order to generalizing the [Expression 6], assuming that the angle between the
projecting plane PL1 and the x axis is and the angle between the projective plane
PL2 and the y axis (axis along a direction orthogonal to the paper plane in Fig. 11) be
a following equations are acquired.

In the above-mentioned equations [Expression 7], (x, y) represents coordinate
values on the image plane PL1, and (x', y") represents coordinate values on the projecting
plane PL2. Parameters "w" and "h" represent the width and the height of the image,

respectively, and are normalized with respect to the width and the height of the image.
This operation corresponds to adjustment of the direction of the optical axis of the
imaging unit.
Subsequently, an image projected onto the projecting plane is subjected to
rotation, scale-changing and translation (parallel displacement). This operation
corresponds to adjustment of the zoom factor of the imaging unit and the rotation of the
optical axis of the imaging unit. Here, provided that the angle of rotation is Go, the
scale-changing rate due to zooming is "k", and the amount of translation is (s', t'),
following expressions are acquired.
[Expression 8]

When the equations [Expression 7] are substituted for the equations [Expression
8], the following equations are acquired.
[Expression 9]

A following relationship is derived by comparing the above described equations
with the projective transformation equations (equations [Expression 1]).
[Expression 10]


Since parameters c, d, t, p and q have already been known in the above
equations [Expression 10], θx, θy, θ0 and k can be determined from these expressions.
In turn, parameters "a" and "b" are determined from these θx, θy, θ0 and k, where "a" and
"b" are determined provided that s = s' = 0.
[StepS132]
When the estimation and calculation of the eight projective transformation
parameters (a, b, s, c, d, t, p and q) is completed according to the above procedure, an
estimation success flag is set, and the estimation of the projective parameters are finished.
In this embodiment, only the pairs of the points satisfying the condition shown
in the equation [Expression 3] are used, and the final parameters are estimated by means
of the least squares method. However, without limitation to this, all the intermediate
parameters can be used to estimate the final parameters.

According to the above estimation method, the projective transformation
parameters for projective-fransforming the feature point are calculated. The acquired
projective transformation parameters are parameters for transforming the reference image.
Intrinsically, the parameters to be calculated are parameters for transforming the target
image. The reason why the parameters for transforming the target image are not
calculated directly will be described below.
Typically, when the inputted image is transformed and outputted, it is calculated
where each pixel position of the outputted image is located on the inputted image. This
is a typical method for avoiding generation of a region without data if it is calculated
where each pixel position of the inputted image is located on the outputted image. In
other words, the parameters required to transform the target image can be parameters for
transforming the feature point in a practical situation.
Although the projective transformation is used as the method for geometrically
transforming the image in this embodiment, a geometrical transformation such as the
Helmert transformation and the affine transformation can be used. In this case, the
estimation method is different from that in the case of the projective transformation to
some extent. For instance, the Helmert transformation is represented as follows:
[Expression 11]
Helmert transformation: X = ax - by + s

Y = bx + ay + t
In the same way as the method for estimating the projective transformation
parameters is adopted, it is applicable to alter calculation of the five parameters (c, d, t, p
and q) to determine Y using the five pairs of the feature point and the corresponding
point for estimating the parameters (a, b and t) to determine Y in the equations
[Expression 11] using three pairs of the feature point and the corresponding point. If
the estimation of the final parameters by the least squares method is performed, it is
applicable to alter the evaluation function J for a following equation.
[Expression 12]

In the Helmert transformation equations, "a" and "b" are included in the
parameters for determining X. Accordingly, if the parameters for determining Y are
estimated, the parameter for determining X is also estimated. The parameter "s", which
is also the translational component with respect to X, can be considered that s = 0, as
with the projective transformation. When applying the geometrical transformation
other than the projective transformation or the Helmert transformation, the parameters
for the geometrical transformation can also be estimated in a similar manner.
The stereoscopic imaging apparatus of this embodiment is the compound eye
camera including the six imaging units. The scope of the presently disclosed subject
matter is not limited to the embodiment. The number of imaging units can be equal to
or larger than two. Furthermore, the stereoscopic imaging apparatus of the presently
disclosed subject matter is limited to the compound eye camera including the plurality of
imaging units. The presently disclosed subject matter is also applicable for a camera
system including a plurality of single eye cameras which are not provided on the one
main body thereof.
Furthermore, the stereoscopic image processing according to the present
invention is not limited to a case of being performed by the stereoscopic imaging
apparatus. Instead, the process can be performed by a personal computer or the like
which does not have a function of taking parallax images. In this case, a plurality of
images (parallax images) taken by a conventional compound eye camera or the like may
be inputted to the personal computer or the like, and the process may be performed on

the personal computer or the like. A function of geometrical transformation for
acquiring the ideal parallax images may be provided as a program which causes the
personal computer to perform the above described process.
The presently disclosed subject matter is not limited to the above-mentioned
embodiment and various modifications can be made without departing from the spirit of
the presently disclosed subject matter.
Reference Signs List
1 to 6...imaging units; 10...stereoscopic imaging apparatus; 12...central
processing unit (CPU); 20...main memory; 22...digital signal processing section;
28...external recording section; 30...displaying section; 32...corresponding point
detecting section; 34...geometrical transformation section; and 36...geometrical
transformation parameter estimation section

CLAIMS
1. A stereoscopic image processing apparatus, comprising:
an image acquisition device for acquiring a plurality of images of an identical
subject taken from a plurality of viewpoints;
a corresponding point detection device for selecting a prescribed image as a
reference image from among the acquired plurality of images, selecting an image other
than the reference image as a target image from among the acquired plurality of images,
and detecting a plurality of feature points from the reference image and a plurality of
corresponding points from the target image to generate a plurality of pairs of the feature
point and corresponding point, wherein feature of the feature point and the corresponding
point included in the same pair are substantially identical to each other;
a parameter estimation device for estimating geometrical transformation
parameters for geometrically transforming the target image such that y coordinate values
of the feature point and the corresponding point included in the same pair are
substantially identical to each other, wherein y direction is orthogonal to a parallax
direction of the plurality of viewpoints; and
an image transformation device for geometrically transforming the target image
on the basis of the estimated geometrical transformation parameters.
2. The stereoscopic image processing apparatus according to claim 1, wherein the
parameter estimation device estimates at least some parameters in the geometrical
transformation parameters based on at least the x and y coordinate values of the
corresponding points and the y coordinate values of the feature points.
3. The stereoscopic image processing apparatus according to claim 2, wherein the
parameter estimation device calculates parameters other than said some parameters
estimated by the parameter estimation device based on said some parameters.
4. The stereoscopic image processing apparatus according to claim 2, wherein
the geometrical transformation parameters are a projective transformation
parameters, and

the parameter estimation device selects five or more pairs from among the
plurality of pairs of the feature point and the corresponding point, and estimates
parameters for determining y coordinate values of projective-transformed corresponding
points on the basis of the coordinate values the feature point and the corresponding point
of each of the selected pairs.
5. The stereoscopic image processing apparatus according to claim 4, wherein the
parameter estimation device calculates the other parameters required to determine the x
coordinate values of the projective-transformed corresponding points on the basis of the
parameters for determining y coordinate values of projective-transformed corresponding
points.
6. The stereoscopic image processing apparatus according to claim 2, wherein
the geometrical transformation parameters are Helmert transformation
parameters, and
the parameter estimation device selects three or more pairs from among the
detected plurality of pairs of the feature point and the corresponding point, and estimates
the Helmert transformation parameters on the basis of the coordinate values of the
feature points and the corresponding points of each of the selected pairs.
7. A stereoscopic image processing method, comprising:
an image acquisition step of acquiring a plurality of images of an identical
subject taken from a plurality of viewpoints;
a corresponding point detection step of selecting a prescribed image as a
reference image from among the acquired plurality of images, selecting an image other
than the reference image as a target image from among the acquired plurality of images,
and detecting a plurality of feature points from the reference image and a plurality of
corresponding points from the target image to generate a plurality of pairs of the feature
point and corresponding point, wherein feature of the feature point and the corresponding
point included in the same pair are substantially identical to each other;
a parameter estimation step of estimating geometrical transformation parameters
for geometrically transforming the target image such that y coordinate values of the

feature points and the corresponding points included in the same pair are substantially
identical to each other, wherein y direction is orthogonal to a parallax direction of the
plurality of viewpoints; and
an image transformation step of geometrically transforming the target image on
the basis of the estimated geometrical transformation parameters.
8. The stereoscopic image processing method according to claim 7, wherein, in the
parameter estimation step, at least some parameters in the geometrical transformation
parameters are estimated based on at least the x and y coordinate values of the
corresponding points and the y coordinate values of the feature points.
9. The stereoscopic image processing method according to claim 8, wherein the
parameter estimation step includes:
a first step of randomly selecting a certain number of pairs required to estimate
the geometrical transformation parameters for determining y coordinate values of the
transformed corresponding points from among N pairs, provided that a total number of
plurality of pairs is N;
a second step of calculating the y coordinate values of the transformed
corresponding points with respect to each of the N pairs based on the parameter
estimated on the basis of the coordinate values of each corresponding point randomly
selected by the first step;
a third step of calculating a difference between the y coordinate value of the
transformed corresponding point calculated by the second step and the y coordinate value
of the feature point for each of the N pairs;
a fourth step of counting a number of pairs of the feature point and the
corresponding point whose difference calculated by the third step is less than a
predetermined first threshold;
a fifth step of determining a confidence level of the estimated parameter on the
basis of a ratio between the counted number of pairs and the N; and
a step of iterating the first to fifth steps until the determined confidence level
reaches a predetermined confidence level, or the number of iterations reaches a
predetermined number of iterations.

10. The stereoscopic image processing method according to claim 9, wherein the
parameter estimation step includes:
a sixth step of calculating the y coordinate value of the transformed
corresponding point for each of the N pairs based on the estimated parameter when the
determined confidence level reaches the predetermined confidence level or the estimated
parameter when the confidence level is highest among the levels at the respective
iterations;
a seventh step of calculating a difference between the y coordinate value of the
transformed corresponding point calculated by the sixth step and the y coordinate value
of the feature point for each of the N pairs;
an eighth step of selecting only the pairs of the feature point and the
corresponding point whose difference calculated by the seventh step is less than a
predetermined second threshold from the N pairs; and
a ninth step of calculating the plurality of parameters using only the pairs of the
feature point and the corresponding point selected in the eighth step.
11. The stereoscopic image processing method according to claim 10, wherein, in
the ninth step, the plurality of parameters which minimize a square sum of differences
between the y coordinate values of the transformed corresponding points in the plurality
of pairs selected by the eighth step and the y coordinate values of the feature points.

12. A stereoscopic imaging apparatus, comprising:
a stereoscopic image processing device according to claim 1; and
a plurality of imaging units which are disposed at a plurality of viewpoints along
the parallax direction respectively, and take images of the identical subject from their
viewpoints, wherein
the image acquisition device acquires the plurality of images taken by the
plurality of imaging units, respectively.

An apparatus (10) includes a device for acquiring a plurali
of images of an identical subject taken from a plurality of viewpoints; a
device for selecting a prescribed image as a reference image, selecting an
image other than the reference image as a target image from among the images,
and detecting feature points from the reference image and corresponding
points from the target image to generate pairs of the feature point
and corresponding point, wherein feature of the feature point and the corresponding
point in the same pair are substantially identical; a device for
estimating geometrical transformation, parameters for geometrically-transforming
the target image such that y-coordinate values of the feature point
and the corresponding point included in the same pair are substantially
identical, wherein y-direction is orthogonal to a parallax direction of the
viewpoints; and a device for geometrically-transforming the target image
based on the parameters.