U.S. patent application number 16/171116 was filed with the patent office on 2019-02-28 for video skimming methods and systems.
The applicant listed for this patent is Vid Scale, Inc.. Invention is credited to Yu Huang, Taoran Lu, Dapeng Oliver Wu, Hong Heather Wu, Zheng Yuan.
Application Number | 20190066732 16/171116 |
Document ID | / |
Family ID | 45556235 |
Filed Date | 2019-02-28 |
![](/patent/app/20190066732/US20190066732A1-20190228-D00000.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00001.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00002.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00003.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00004.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00005.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00006.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00007.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00008.png)
![](/patent/app/20190066732/US20190066732A1-20190228-D00009.png)
![](/patent/app/20190066732/US20190066732A1-20190228-M00001.png)
View All Diagrams
United States Patent
Application |
20190066732 |
Kind Code |
A1 |
Lu; Taoran ; et al. |
February 28, 2019 |
Video Skimming Methods and Systems
Abstract
In an embodiment, an apparatus and method of creating a skimming
preview of a video includes electronically receiving a plurality of
video shots, analyzing each frame in a video shot from the
plurality of video shots, where analyzing includes determining a
saliency of each frame of the video shot. The method also includes
determining a key frame of the video shot based on the saliency of
each frame the video shot, extracting visual features from the key
frame, performing shot clustering of the plurality of video shots
to determine concept patterns based on the visual features, and
generating a reconstruction reference tree based on the shot
clustering. The reconstruction reference tree includes video shots
categorized according to each concept pattern.
Inventors: |
Lu; Taoran; (Santa Clara,
CA) ; Yuan; Zheng; (San Jose, CA) ; Huang;
Yu; (Sunnyvale, CA) ; Wu; Dapeng Oliver;
(Gainesville, FL) ; Wu; Hong Heather; (West
Windsor, NJ) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Vid Scale, Inc. |
Wilmington |
DE |
US |
|
|
Family ID: |
45556235 |
Appl. No.: |
16/171116 |
Filed: |
October 25, 2018 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
14922936 |
Oct 26, 2015 |
10153001 |
|
|
16171116 |
|
|
|
|
13103810 |
May 9, 2011 |
9171578 |
|
|
14922936 |
|
|
|
|
61371458 |
Aug 6, 2010 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06K 9/00751 20130101;
G11B 27/034 20130101; G11B 27/031 20130101; G11B 27/28
20130101 |
International
Class: |
G11B 27/031 20060101
G11B027/031; G11B 27/034 20060101 G11B027/034; G11B 27/28 20060101
G11B027/28; G06K 9/00 20060101 G06K009/00 |
Claims
1. An apparatus comprising: a processor; a memory coupled to the
processor; a port coupled to the processor to electronically
receive a plurality of video shots; and a non-transitory
computer-readable medium storing instructions that are operative,
when executed by the processor to perform acts including: analyzing
each frame in a video shot from the plurality of video shots, the
analyzing comprising determining a saliency of each frame of the
video shot, the saliency being a content attentiveness saliency
providing a measurement of representative shot properties;
determining an effective visual saliency based on the determined
saliency, the effective visual saliency based on camera motion of
the plurality of video shots and a human attention model of the
video shot; selecting a key frame of the video shot based on the
effective visual saliency of each frame the video shot; extracting
visual features from the key frame; performing shot clustering of
the plurality of video shots to determine concept patterns based on
the visual features; and generating a hierarchical reconstruction
based on the shot clustering, the hierarchical reconstruction
enabling a skimming preview of the video.
2. The apparatus of claim 1, wherein the human attention model is
based on camera motion.
3. The apparatus of claim 2, wherein: the human attention model
comprises a camera attenuation factor based on the camera motion,
wherein the processor to performs acts further including:
determining the camera attenuation factor; and determining the
effective visual saliency comprises multiplying the determined
effective visual saliency with the determined camera attenuation
factor.
4. The apparatus of claim 3, wherein the determined camera
attenuation factor is proportional to a zooming speed of video shot
from the plurality of video shots.
5. The apparatus of claim 3, wherein the determined camera
attenuation factor is inversely proportional to a panning speed of
video shot from the plurality of video shots.
6. The apparatus of claim 1, wherein the hierarchical
reconstruction comprises a reconstruction reference tree that
includes video shots categorized according to each concept
pattern.
7. The apparatus of claim 6, wherein generating the reconstruction
reference tree comprises categorizing video shots within concept
categories ordered according to concept importance, and ordering
video shots within each concept category according to effective
visual saliency.
8. The apparatus of claim 7, wherein the concept importance
includes determining a total number of frames in a shot having a
same concept.
9. The apparatus of claim 6, further comprising wherein the
hierarchical reconstruction includes generating a video skimming
preview based on the reconstruction reference tree.
10. The apparatus of claim 1, wherein the processor performs acts
further including extracting audio features from the video
shot.
11. The apparatus of claim 10, wherein extracting audio features
comprises: determining audio words from the video shot; and
performing clustering on the audio words.
12. The apparatus of claim 11, further comprising: determining
visual concept patterns based on the performing shot clustering;
and determining audio concept patterns based on the performing
clustering on the audio words.
13. The apparatus of claim 12, further comprising: calculating a
number of member shots for each visual concept of the visual
concept patterns and for each audio concept of the audio concept
patterns; sorting each visual concept by calculated number of
shots; sorting each audio concept by calculated number of shots;
and aligning visual concepts and audio concepts having a same
number of shots.
14. A method comprising: electronically receiving a reconstruction
reference tree comprising video shots categorized within concept
categories ordered according to concept importance, wherein video
shots within each concept category is ordered according to
saliency; selecting shots starting from categories of highest
importance and shots of highest saliency within the categories of
highest importance; and generating a preview based on the selected
shots.
15. The method of claim 14, further comprising: selecting frames
within the selected shots having a highest saliency; and using the
frames within the selected shots having the highest saliency to
generate the preview.
16. The method of claim 15, wherein selecting frames within the
selected shots having a highest saliency comprises: selecting a
target skimming ratio; determining a threshold according to the
selected target skimming ratio; and comparing the saliency of the
selected shots to the threshold.
17. The method of claim 16, wherein the selected target skilling
ratio is an arbitrary length.
18. The method of claim 14, wherein the saliency comprises an
effective visual saliency based on camera motion of the video
shots.
19. A non-transitory computer readable medium with an executable
program stored thereon, wherein the program instructs a
microprocessor to perform the following steps: analyzing each frame
in a video shot from a plurality of video shots, the analyzing
including determining a saliency of each frame of the video shot,
the saliency being a content attentiveness saliency; determining an
effective visual saliency based on the determined saliency and
based on camera motion of each from the video shot; selecting a key
frame of the video shot based on the effective visual saliency of
each frame of the video shot; extracting visual features from the
key frame; performing shot clustering of the plurality of video
shots to determine concept patterns based on the visual features;
and generate a reconstruction reference tree based on the shot
clustering, the reconstruction reference tree comprising video
shots categorized according to each concept pattern.
20. The non-transitory computer readable medium of claim 19,
wherein the program instructs the microprocessor to further perform
the steps of: determining audio features of the video shot;
determining saliency of the determined audio features; clustering
determined audio features; and aligning audio and video concept
categories.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application is a continuation application of U.S.
patent application Ser. No. 14/922,936, filed Oct. 26, 2015,
entitled "Video Skimming Methods and Systems," which is a
continuation of U.S. patent application Ser. No. 13/103,810, filed
May 9, 2011, now issued as U.S. Pat. No. 9,171,578, which claims
priority from U.S. Provisional Application, Ser. No. 61/371,458,
filed Aug. 6, 2010, entitled "Video Skimming Methods and Systems,"
which application is hereby incorporated herein by reference in its
entirety.
TECHNICAL FIELD
[0002] The present invention relates to image processing, and, in
particular embodiments, to video skimming by the perspective of
hierarchical audio-visual reconstruction with saliency-masked
Bag-of-Words features.
BACKGROUND
[0003] The fast evolution of digital video has brought many new
applications and consequently, research and development of new
technologies, which will lower the costs of video archiving,
cataloging and indexing, as well as improve the efficiency,
usability and accessibility of stored videos arc greatly needed
Among all possible research areas, one important topic is how to
enable a quick browse of a large collection of video data and how
to achieve efficient content access and representation.
[0004] To address these issues, video abstraction techniques have
emerged and have been attracting more research interest in recent
years There are two types of video abstraction, video summary and
video skimming Video summary, also called a still abstract, is a
set of salient images (key frames) selected or reconstructed from
an original video sequence.
[0005] Video skimming, also called a moving abstract, is a
collection of image sequences along with the corresponding audios
from an original video sequence Video skimming is also called a
preview of an original video, and can be classified into two
sub-types: highlight and summary sequence A highlight contains the
most interesting and attractive parts of a video, while a summary
sequence renders the impression of the content of an entire video
Among all types of video abstractions, summary sequence conveys the
highest semantic meaning of the content of an original video.
[0006] One prior art method is uniform sampling the frames to
shrink the video size while losing the audio part, which is similar
to the fast forward function seen in many in digital video players.
Time compression methods can compress audio and video at the same
time to synchronize them, using frame dropping and audio sampling.
However, the compression ratio can be limited by speech distortion
in some cases. Frame-level skimming mainly relies on the user
attention model to compute a saliency curve, but this method is
weak in keeping the video structure, especially for a long video.
Shot clustering is a middle-level method in video abstraction, but
its readability is mostly ignored. Semantic level skimming is a
method trying to understand the video content, but can be difficult
to realize its goal due to the "semantic gap" puzzle.
SUMMARY OF THE INVENTION
[0007] In accordance with an embodiment, a method of creating a
skimming preview of a video includes electronically receiving a
plurality of video shots, analyzing each frame in a video shot from
the plurality of video shots, where analyzing includes determining
a saliency of each frame of the video shot. The method also
includes determining a key frame of the video shot based on the
saliency of each frame the video shot, extracting visual features
from the key frame, performing shot clustering of the plurality of
video shots to determine concept patterns based on the visual
features, and generating a reconstruction reference tree based on
the shot clustering. The reconstruction reference tree includes
video shots categorized according to each concept pattern.
[0008] The foregoing has outlined rather broadly the features of an
embodiment of the present invention in order that the detailed
description of the invention that follows may be better understood.
Additional features and advantages of embodiments of the invention
will be described hereinafter, which form the subject of the claims
of the invention. It should be appreciated by those skilled in the
art that the conception and specific embodiments disclosed may be
readily utilized as a basis for modifying or designing other
structures or processes for carrying out the same purposes of the
present invention. It should also be realized by those skilled in
the art that such equivalent constructions do not depart from the
spirit and scope of the invention as set forth in the appended
claims,
BRIEF DESCRIPTION OF THE DRAWINGS
[0009] For a more complete understanding of the present invention,
and the advantages thereof, reference is now made to the following
descriptions taken in conjunction with the accompanying drawing, in
which:
[0010] FIG. 1 provides frame shots illustrating embodiment visual
saliency masking;
[0011] FIG. 2 illustrates embodiment SIFT features in active
regions on video frames;
[0012] FIG. 3 provides a graph to illustrating an embodiment BoW
feature for example video frames;
[0013] FIG. 4 provides a flow chart of an embodiment visual BoW
feature extraction;
[0014] FIG. 5 illustrates an embodiment saliency masking of audio
words;
[0015] FIG. 6 illustrates a flow chart of audio BoW feature
extraction;
[0016] FIG. 7 illustrates an example graph of an audio BoW
feature;
[0017] FIG. 8 illustrates an embodiment reconstruction reference
tree and shot table;
[0018] FIG. 9 illustrates an embodiment post processing saliency
curve;
[0019] FIG. 10 illustrates and embodiment system that implements
embodiment algorithms;
[0020] FIG. 11 illustrates a processing system that can be utilized
to implement embodiment methods; and
[0021] FIG. 12 provides an appendix listing notations used
herein.
DETAILED DESCRIPTION OF ILLUSTRATIVE EMBODIMENTS
[0022] The making and using of the presently preferred embodiments
are discussed in detail below It should be appreciated, however,
that the present invention provides many applicable inventive
concepts that can be embodied in a wide variety of specific
contexts The specific embodiments discussed are merely illustrative
of specific ways to make and use the invention, and do not limit
the scope of the invention.
[0023] A novel approach to video summarization is disclosed This
approach includes unsupervised learning of original video concepts
and hierarchical (both frame and shot levels) reconstruction.
[0024] Video skimming is a task that shortens video into a
temporally condensed version by which viewers may still understand
the plot of original video. This technique allows viewers to
quickly browse of a large multimedia library and thus facilitates
tasks such as fast video browsing, video indexing and retrieval.
The performance of video summarization mainly lies in the following
three aspects: skeleton preservation, exciting and interesting
summarization and smooth transition.
[0025] Video summarization enables viewers to quickly and
efficiently grasp what a video describes or presents from a shorter
summarized version. To meet this need, the main skeleton from the
original video is extracted and kept it in the summarized video. A
video skeleton can be seen as a queue of concept patterns with
certain semantic implications in a temporal order. The term
"Concept pattern" here is not a high-level as real semantic concept
that is learned by human intervention. Rather, a concept pattern
encodes semantic meanings of shots (sets of consecutive similar
video frames), symbolizes a shot group that portraits consistent
semantic settings, and generally possesses the capability as a
hallmark or self-evident clue that hints to the development of the
original video. Viewers may possibly recover the plot by only
watching and hearing a handful of shots as long as all concept
patterns are conveyed.
[0026] Some embodiments of the present invention present to viewers
an exciting interesting summary of video. Often in a video, there
are various shots conveying the same concept patterns. When
selecting one shot conveying a concept pattern from many, the one
with high saliency value or equivalently generating the largest
stimulus to human attention would be favored so that the resultant
summarized video not only contains integral concept patterns, but
also carefully selects shot instances with richest information to
reflect these concept patterns. Hence, a plain or even dull
summarization is avoided.
[0027] In some cases, an unnatural transition between two adjacent
concept patterns due to the elimination of a number of visually and
acoustically similar shots is apparent in conventional video
skimming previews. Some embodiments of the present invention
provide a smooth transition between adjacent concept patterns by
providing additional frame level summarization.
[0028] Embodiments of the present invention generate video
summarization by providing unsupervised learning of original video
concepts and hierarchical (both frame and shot levels)
reconstruction. In an embodiment, the skeleton of original video is
analyzed by concept pattern mining. Viewing it as a clustering
problem. Bag of Words features (SIFT based visual word and
Matching-pursuit based audio word) are extracted for each shot from
both visual and audio sensory channels that filtered with saliency
masking. The Bag of Words features are then clustered into several
groups using spectral clustering techniques. Each cluster
represents a certain concept pattern.
[0029] Next, based on the discovered concept patterns, the original
video is summarized from reconstruction point of view based on the
learned concept pattern. In some embodiments, summarization is
regarded as a "summation" process rather than a "subtraction"
process. Keeping at least one shot for each concept pattern, the
concept pattern integrity of summarized video offer viewers the
capability of context recovery. In addition, given a specified
skimming ratio, a video that also contains maximum achievable
saliency accumulation is generated. In some embodiments, the
summarization process is conducted in an iterative fashion,
allowing flexible control of summarized video information richness
vs. skimming ratio.
[0030] Finally, to meet the skimming ratio specification and keep
the smooth transition in the summarized video, a frame level
saliency thresholding is used, which is followed by a temporally
morphological operation as post processing.
[0031] A good understanding of video content can help achieve a
good video summary. The most common contents for a typical video
sequence are visual and acoustic channels. Most of the time, visual
signals provide the majority information from which latent concept
patterns are learned from original video. However, audio sensory
channels can also provide important information regarding concept
pattern in ways not offered by the visual channel, for example, in
low light environments and nighttime shots. Furthermore, in
embodiments, a concept pattern can be derived that also shares both
visual and audio consistency at the same time. Thus, if independent
feature extraction and unsupervised concept learning from both
visual and audio sensory data is used, learned concept results can
be jointly analyzed in a parity-check fashion to enhance
co-reliability. Therefore, in some embodiments, an audio stream is
extracted from raw video and processed in parallel with the video
stream to detect possible audio concepts.
[0032] In an embodiment, temporal segmentation for video stream is
archived using shot detection. A variance-difference based approach
is used to detect a shot change, and robustly detects scene cuts
and fades between scenes. The variance of a frame is calculated and
the delta variance with its previous frame Dvar is recoded. In an
embodiment, the criteria for Dvar to start a new shot are:
[0033] a. Dvar (current)<Th1 (stability requirement)
[0034] b. maxDvar(start to current)-minDvar(start to
current)>Th2 (tolerance requirement)
[0035] c. Frame number in current shot>Th3 (shot length
requirement)
[0036] In alternative embodiments, other techniques can be used.
For example, shot boundaries can also be found using color
histogram based approaches or optical-flow motion features. For
processing convenience, in some embodiments, audio data are
segmented into pieces, where each piece has its boundaries
synchronized to its co-located video shot in time axis.
[0037] In an embodiment, an attention model and a bag-of-words
feature construction on shots is performed. Skeleton preservation
uses some distinctive feature for shot-discrimination, and an
exciting summarization uses a content attentiveness (saliency)
measurement. Embodiment saliency measurement methods effectively
reflect how informative a shot is, and shot features are selected
to represent video skeleton with discrimination (i.e., to be used
to find the similarity with other shots).
[0038] In embodiments, Bag-of-Words (BoW) models are used to
characterize the shot properties in visual and aural domains,
respectively. The Bag-of-Words (BoW) model was initially utilized
in Natural Language Processing to represent the structure of a
text. For example, a BoW model regards a text document as a
collection of certain words belonging to a reference dictionary but
ignores the order and semantic implications of words. A BoW model
uses the occurrence of each word in the dictionary as the feature
of text, and therefore, often produces a sparse vector. The BoW
model can be regarded as the "histogram representation based on
independent features." As such, a shot can be regarded as a sort of
"text document" with regard to some embodiments. However, since
neither the "visual word" nor the "aural word" in a shot is readily
apparent like real words in a text document, such visual and aural
"words" need to be defined. In an embodiment, the determination of
a "word" usually involves two steps: feature extraction and
codeword generation.
[0039] In an embodiment, visual words are derived using saliency
detection according to PQFT-based attention modeling. Such an
attention model has been shown to be successful in imitating
human's perceptual properties on video frames. The generated
saliency map is used as a good indicator of how conspicuous a frame
is, and which part within the frame incurs the highest human
attention. Given the saliency map for each frame, a measure of
visual frame-saliency is formulated by calculating the average
value of the saliency map for a frame t:
Sal t v = 1 W .times. H i = 1 W j = 1 H WM t ( i , j ) ,
##EQU00001##
where W and H are frame width and height respectively, SM refers to
the saliency map for frame t.
[0040] For a shot, the visual conspicuous level is calculated by
averaging the visual frame saliency in that shot:
AvgSal k v = { 1 N k t Sal t v | F t .di-elect cons. Shot k }
##EQU00002##
[0041] In an embodiment, the visual structure of original video
observed from a middle-level video concept pattern is derived. In
general, a video concept pattern can be viewed as a self-learned
set featured by a combination of certain Spatially Local Visual
Atom (SLVA) and each SLVA stands for a single visual pattern, which
is found within a localized neighborhood at a particular spatial
location, with plausible semantic implications, like green grass,
blue sky, etc. A noticeable property of the video concept pattern
is that, importance is only attached to the occurrence of SLVAs,
without esteeming the order (spatial location). For example, a shot
of a far view of green hills with blooming yellow flowers and a
shot of a near view of the grass and flowers should both imply the
same concept, even though the grass and flowers may appear in
different locations and in different scales. As such, the BoW model
for visual shots, which graciously expresses the order-irrelevant
property, is employed by embodiment of the present invention using
SLVAs as the visual words. Alternatively, other techniques, such as
part-based methods, referred in B Leibe, A Leonardis, and B
Schiele, "Robust Object Detection with Interleaved Categorization
and Segmentation", IJCV Special Issue on Learning for Vision and
Vision for Learning, August 2007, can be used.
[0042] In a regular full process mode, SIFT feature points are
generally detected on every frame in the shot and on every region
within a frame. This procedure, although precise, is especially
time-consuming. Thus, some embodiments employ pre-processing steps
prior to SIFT feature detection. In an embodiment key frames are
used to balance computational cost and accuracy. Further, a
filtering process called saliency masking is used to improve the
robustness and efficiency of the SIFT feature extraction in some
embodiments.
[0043] Considering the fact that some frames within a shot appear
to have minor differences, one frame, referred to as a key frame,
is selected as the most representative frame in a shot. There are
many key frame selection methods known in the art that can be used.
Some straightforward methods include choosing the first/last frame,
or the middle frame in a shot. Some motion-based approaches use
motion intensity to guide the key frame selection, such as those
used in MEPG-7. In an embodiment, however, human attention models
are used, and a most salient frame is used to represent a shot as
follows:
t.sub.k=arg max(Sal.sub.t.sup.v|F.sub.t Shot.sub.k)
[0044] Embodiment key frame selection techniques can save a large
amount of computation resources at a minor cost of precision loss,
assuming that frames are similar within a shot. If such an
assumption does not hold, the attention model can be exploited with
respect to a single frame to exclude some inattentive regions on
the key frame. An embodiment attention model, based on image phase
spectrum and motion residue is used to imitate human perceptual
properties. In an embodiment, a saliency map SM is generated whose
pixel value indicates how attentive the pixel on original frame
is.
[0045] The movement of the camera through an environment, e.g., a
fixed background is the ego-motion. In an embodiment, the impact of
ego-motion is incorporated in determining the saliency of a frame.
As described further in detail, this is accomplished by computing a
motion channel having a difference in intensity map and an
ego-motion compensated intensity map.
[0046] In an embodiment, the camera motion between adjacent frames
is. Camera motion between two adjacent frames can be computed by
estimating a 2-D rigid transformation based on the corresponding
KLT (Kanade-Lucas-Tomasi Feature Tracker) key point tracks on the
two adjacent frames. Embodiments of the invention also include
alternative methods such as SIFT matching or Speeded Up Robust
Features (SURF) correspondence etc.
[0047] Suppose a KLT key point is located at (x,y) in frame t, the
corresponding KLT key point is tracked at (x',y') in frame (t+1),
and the transformation from (x,y) to (x',y') can be expressed as
follows:
[ x ' y ' ] = A [ x y ] + b = [ s cos .theta. s sin .theta. - s sin
.theta. s cos .theta. ] [ x y ] + [ b x b y ] . ##EQU00003##
[0048] In the above equation, s, .theta., b, and by are camera
parameters, wherein s is the zooming factor, .theta. is the
counterclockwise rotation degree, b.sub.x corresponds to the pan
movement, and b.sub.y corresponds to the tilt movement.
[0049] For a set of KLT key point correspondences, matrix A and
vector b may be solved using the robust RANSAC (RANdom SAmple
Consensus) rigid transformation estimation, which is a known
iterative method to estimate parameters of a mathematical model
from a set of observed data having outliers. RANSAC is a
non-deterministic algorithm in the sense that it produces a
reasonable result only with a certain probability, which increases
with the number of allowed iterations. Embodiments of the invention
may also use alternative methods such as Least Median of Squares or
M-Estimator etc.
[0050] After estimating the camera motion parameters, the visual
saliency of each frame may be determined. The camera motion may be
applied to compensate the ego motion and the residual may be fused
into the color information to generate visual saliency.
[0051] Next, the intensity channel I(t) of a frame t is calculated
using the color channels of the frame as follows. A given frame t
may comprise red r(t), green g(t), and blue b(t) channels. Four
broadly tuned color channels may be generated by the following
equations:
R(t)=r(t)-(g(t)+b(t))/2
G(t)=g(t)-(r(t)+b(t))/2
B(t)=b(t)-(r(t)+g(t))/2
Y(t)=(r(t)+g(t))/2-|r(t)-g(t)|/2-b(t).
[0052] In addition, two color difference channels are defined as
following.
RG(t)=R(t)-G(t)
BY(t)=B(t)-Y(t).
[0053] The intensity channel is calculated as follows:
I(t)=(r(t)+g(t)+b(t))/3.
[0054] The ego-motion compensated intensity map I(t-.tau.) for the
previous frame (t-.tau.) is computed. The motion channel M(t) is
computed as an absolute difference between intensity map I(t) and
ego-motion compensated intensity map I(t-.tau.) as follows:
M(t)=|I(t)-(A.sub.t-.tau..sup.t(t-.tau.)+b.sub.t-.tau.T.sup.t)|.
[0055] In the above equation, A.sub.t-.tau..sup.t and
b.sub.t-.tau..sup.t t are the estimated camera parameters from
frame (t-.tau.) to frame t.
[0056] Next, the frame t can be represented as a quaternion image
q(t):
q(t)=M(t)+RG(t).mu..sub.1+BY(t).mu..sub.2+I(t).mu..sub.3.
[0057] In the above equation, .mu..sub.1.sup.2=-1, j=1, 2, 3; and
.mu..sub.1.perp..mu..sub.2, .mu..sub.1.perp..mu..sub.3,
.mu..sub.2.perp..mu..sub.3, .mu..sub.3=.mu..sub.1.mu..sub.2.
[0058] Furthermore, q(t) can be represented in symplectic form as
follows:
q(t)=(t)+f.sub.2(t).mu..sub.2
(t)=M(t)+RG(t).theta..sub.1
f.sub.2(t)=BY(t)I(t).mu..sub.1.
[0059] A Quaternion Fourier Transform (QFT) is performed on the
quaternion image q(n,m,t), where (n,m) is the location of each
pixel in time domain:
Q[u, v]=F.sub.1[u, v]+F.sub.2[u, v].mu..sub.2
F i [ u , v ] = 1 MN m = 0 M - 1 n = 0 N - 1 e - .mu. 1 2 .pi. ( (
mv / M ) + ( nu / N ) ) f i ( n , m ) . ##EQU00004##
[0060] In the above equations, (u,v) is the location of each pixel
in frequency domain, while N and M are the image's height and
width.
[0061] The inverse Fourier transform is obtained as follows:
f i [ n , m ] = 1 MN m = 0 M - 1 n = 0 N - 1 e .mu. 1 2 .pi. ( ( mv
/ M ) + ( nu / N ) ) F i [ u , v ] . ##EQU00005##
[0062] A Frequency domain representation Q(t) of the quaternion
image q(t) can be rewritten in the polar form as follows:
Q(t)=||Q(t)||e.sup..mu..phi.(t),
where .phi.W is the phase spectrum of Q(t).
[0063] In equation 20, if ||Q(t)||=1, the frequency domain
representation Q(t) of the quaternion image q(t) includes only the
phase spectrum in frequency domain. Therefore, the inverse
Quaternion Fourier Transform (IQFT) of the phase spectrum of the
frequency domain representation Q(t) of the quaternion image q(t)
may be performed. The IQFT of the phase spectrum q'(t) is a 2-D
image map and may be computed as follows:
q40 )t)=a(t)+b(t).mu..sub.1+c(t).mu..sub.2+d(t).mu..sub.3.
[0064] The saliency map (sM(t)) of frame t may be obtained by
taking a smoothing filter kernel and running a convolution with the
2-D image map q'(t):
sM(t)=g*||q'(t)||.sup.2,
[0065] where g is a 2-D Gaussian smoothing filter. In various
embodiments, for computation efficiency, only the Phase Spectrum of
Quaternion Fourier Transform (PQFT) on a resized image (e.g., whose
width equals to 128) may be computed.
[0066] Next, the visual saliency value S.sub.v(t) of the frame t
may be computed by taking the average over the entire saliency map
as follows:
S v ( t ) = 1 MN m = 0 M - 1 n = 0 N - 1 sM ( n , m , t ) .
##EQU00006##
[0067] Embodiments of the invention for tuning the saliency to
account for camera motion will next be described. Camera motion may
be utilized to emphasize or neglect certain objects. Alternatively,
camera motion may be used to guide viewers' attentions during a
scene.
[0068] In one or more embodiments, the rigid motion estimation as
described above, may be used to determine the camera motion type
and speed. However, further information is required to understand
the relationship between camera motion and the significance of a
particular camera motion in guiding a user. For example, it is
necessary to be able to map the computed camera parameters to their
ability to attract a viewer's attention. Embodiments of the
invention use general camera work rules to set up a user attention
based model.
[0069] The user attention based model is obtained based on the
following assumptions from general movie production. First, zooming
is assumed to emphasize something. In particular, the speed of
zooming scales linearly with the importance of the media segment.
Therefore, faster zooming speeds describe important content.
Usually, zoom-in is used to emphasize details, while zoom-out is
used to emphasize an overview scene. Second, a video producer may
apply panning if the video producer wants to neglect or
de-emphasize something. As in zooming, the speed of the panning
operation may be used a metric of importance. Unlike zooming, the
faster the panning speed is, the less important the content is.
[0070] [The visual saliency value S.sub.v(t) of frame t is then
scaled by the corresponding camera attention factor
.omega..sub.cm(t). Therefore, the effective visual saliency
S.sub.v*(t) is computed as:
S.sub.v*(t).rarw..omega..sub.cm(t)S.sub.v(t).
[0071] In various embodiments, an attention factor Worn caused by
camera motion is quantified over a pre-determined range, for
example, [0{tilde over ( )}2]. For example, a value greater than 1
may represent emphasis, while a value smaller than 1 may represent
neglect.
[0072] Next, an active region on the key frame is defined by
thresholding the saliency map:
AR.sub.t k(i,j)={F.sub.t k(i,j)|SM.sub.t k(i,j)>T,
1.ltoreq.i.ltoreq.W,1.ltoreq.j.ltoreq.H}
where, T is the active threshold. The SIFT feature detection in
remaining active regions then generates prominent and robust SLVAs
of the frame.
[0073] FIG. 1 illustrates the results of saliency masking on shot
19 and 45 of a frame sequence. For example, frame 102 represents
shot 19 prior to the application of saliency masking, and frame 104
represents shot 19 after saliency masking has been applied. Regions
103 represent the masked regions of shot 19. Similarly, frame 106
represents shot 45 prior to the application of saliency masking,
and frame 108 represents shot 45 after saliency masking has been
applied. Regions 105 represent the masked regions of shot 45.
[0074] In an embodiment, Lowe's algorithm for SIFT feature
detection in active regions on the key frame is used. The frame is
convolved with Gaussian filters at different scales, and then the
differences of successive Gaussian-blurred versions are taken. Key
points are located as maxima/minima of the Difference of Gaussians
(DoG) that occur at multiple scales. Then, low-contrast key points
are discarded and high edge responses are eliminated. Next, each
key point is assigned one or more orientations based on the local
gradient directions. Finally, a highly distinctive 128-dimension
vector is generated as the point descriptor; i.e., the SLVA. For
example, FIG. 2 shows detected SIFT feature points 109 in frames
104 and 108 representing shots 19 and 45, respectively.
[0075] After SIFT feature points are found on the key frame of each
shot, the shot as a bag has a collection of "visual-word," with
each is a vector of dimension 128. The number of words is the
number of SIFT feature point s on the key frame. A shot bag with
its SIFT feature descriptors can now be regarded as a text document
that has many words. In order to generate the histogram
representation as the feature for the shot, "dictionary" is built
as the collection of all the "words" from all the bags, and similar
"words" should be treat as one "codeword," as in text documents,
"take", "takes", "taken" and "took" should be regarded
same--"take", as its codeword. The bags of words in visual
appearance are referred to in L. Fei-Fei and P. Perona, "A Bayesian
Hierarchical Model for Learning Natural Scene Categories," by IEEE
Computer Vision and Pattern Recognition. pp. 524-531, 2005, which
is incorporated herein by reference. Alternatively, other
algorithms can be used, such as, but not limited to those described
in G. Csurka, C. Dance, L. X. Fan, J. Willamowski, and C. Bray.
"Visual categorization with bags of keypoints". Proc. of ECCV
International Workshop on Statistical Learning in Computer Vision,
2004, can be used. Furthermore, other vector dimensions can be used
as well.
[0076] A codeword can be considered as a representative of several
similar SLVAs. In an embodiment, K-means clustering over all the
SLVAs is used, where the number of the clusters is the codebook
size. Such an embodiment can be viewed as being analogous to the
number of different words in a text dictionary. Here, codewords are
the centers of the clusters, and each "word" is mapped to a certain
codeword through the clustering process.
[0077] Thus, each shot can be represented by a histogram of the
codewords. In one example, to take into account the complexity of a
particular video sequence, 200 codewords are used. FIG. 3 depicts
the visual BoW feature for 19, frame 104 shown in FIGS. 1 and 2
above. Here, the x-axis represents the index of words, and the
y-axis represents the normalized frequency of words occurred in the
key frame of the shot.
[0078] FIG. 4 illustrates flowchart 200 of an embodiment visual BoW
feature extraction for a shot. Key frame 202 of SHOT.sub.k is input
to SIFT feature detection block 204, which applies SIFT feature
detection to SHOT.sub.k. Each detected SLVA 206 is assigned a
codeword in step 208 based on codeword generation by K-means
clustering block 210. The frequency of each determined codeword 218
is counted in step 220 to produce visual BoW 222 for SHOT.sub.k.
Codeword generation by K-means clustering block 210 generates
codewords based on all SLVAs (block 212) found among the key frame
of SHOT.sub.k 202, as well as and key frames 216 and SLVAs 214 from
other shots. Visual BoWs 224 for other shots are similarly
derived.
[0079] In an embodiment, the audio structure of the original video
is observed with respect to an audio concept pattern. In general,
an audio concept pattern is interpreted as acoustic environment
featured by a combination of certain Temporally Local Acoustic Atom
(TLAA). Each TLAA stands for a single audio pattern with plausible
semantic implications. For example, an audio concept conversation
between John and Mary at the shore is featured as a combination of
John's short time voice (a TLAA) switching with Mary's (a TLAA) and
continuous environmental sound of sea wave (a TLAA). Note that for
the purpose of video summarization, an audio skeleton is sought
that are usually comprised of "self-contained" concept patterns,
meaning that in the set of shots that form a concept pattern, every
shot has TLAAs from the same closed subset of plausible audio
patterns and the reshuffling of plausible audio patterns is
allowed. This assumption originates from the fact that humans
recognize an audio scene from a macroscopic perspective, which
emphasizes the components instead of an exact time and location of
every component.
[0080] As in the above example, if another audio scene also
includes John, Mary and the sea wave, but this time John
continuously talks during the first half and Mary talks during the
second half, without any voice switching, this scene is still
considered to have the same concept pattern as the example above.
Here, the second example also conveys the semantic implication of
John Mary's conversation at the shore. With respect to one audio
concept, those shots are subject to consistent TLAA compositions,
regardless of which order these TLAAs are arranged.
[0081] In the context of audio concept clustering, at this level,
the feature vectors of different shots may be much closer to each
other, as long as their acoustic component TLAAs are alike. In some
embodiments, they are then pruned to be clustered into the same
group, which captures the underlying common characteristics of an
audio scene. Compared to many indicator-like features, which
identifies a shot as a single acoustic source, for example, speech
from a single person, sound from a machine or environment, and
background music, each shot will end up to be a sparse vector with
only one 1-entry that indicates to which acoustic source this shot
belongs. While this hard-decision-like feature can be viewed as
contradictory to the fact that an audio segment corresponding to a
shot usually consists of multiple intervening sources, this fact is
implicitly reflected by a BoW feature.
[0082] For indicator-like features, the sparse nature of their shot
data highlights the difference of shot data by assuming shot as a
single source with majority contribution, which are usually
different. In this way, the clustering may lose much opportunity to
learn a reasonable concept pattern where shots have similar
acoustic components, but the majority of sources are different.
Therefore, embodiment BoW features encode intervening sources of a
concept softly, which provides a closer approximation to the nature
of an underlying concept as perceived by humans, and thus yields
more accuracy.
[0083] To serve the needs of concept pattern mining that focuses on
components rather than their order, the BoW model can suitably
represent the audio features of a detected shot. If the audio
stream of a shot is chopped into multiple overlapped short-time
audio segments with equal length, the shot can be regarded as a bag
containing multiple audio segments as audio words. Each word, with
extracted feature by Matching Pursuit decomposition, represents a
unique TLAA, which is an audio pattern with plausible semantic
implications. A shot is consequently considered as a bag containing
the audio patterns. The histogram of each word occurrence is a
summarized feature of a shot through all the words within. Here, an
encoding theme is applied to avoid the over-sparsity of feature
vectors (negatively impact the classification result) from a direct
word occurrence statistic. In an embodiment, all audio words from
all shots in raw video are stored into a dictionary, and K-means
clustering is conducted over the dictionary to produce K codewords.
Each word is then assigned to a nearest codeword. The BoW feature
of each shot is the occurrence of codewords inside the shot.
[0084] In an embodiment, the robustness of an audio BoW feature is
improved by taking into account audio words above an acoustic
saliency level to avoid the negative effect on the BoW accuracy
exerted by audio words of low saliency. This can be due to its
small value compared with noise. Here, audio saliency is measured
by a variety of low-level audio features (scalar values), including
Spectral Centroid, Root Mean Square (RMS), Absolute Value Maximum,
Zero-Crossing Ratio (ZCR), and Spectral Flux. By using saliency
masking, the audio words experience a reliability test so that the
accuracy of features for every word is increased.
[0085] The spectral centroid is the center of the spectrum of the
signal, and is computed by considering the spectrum as a
distribution whose values are the frequencies, and the
probabilities to observe these are the normalized amplitude. Root
mean square is a measure of short time energy of a signal from norm
2. Absolute Value Maximum is a measure of short time energy of a
signal from norm 1. The zero-crossing is a measure of the number of
time the signal value cross the zero axis. These measures are
further discussed by G. Peeters. "A large set of audio features for
sound description (similarity and classification) in the CUIDADO
project,"Report for the Institute De Recherche Et Coordination
Acoustique/Musique, April 2004, which is incorporated herein by
reference. If the numbers of members of more than one concept are
equal, then match ambiguity emerges in the one-to-one mapping
between visual and audio concepts. Here the time signature of shot
member is used as alternative feature to bind two concept sets. A
concept pair producing more similar time signature on both sides
are considered as a good pair and are matched up. The time
signature is the starting/ending time and duration of a shot.
[0086] In an embodiment, an algorithm is used to progressively
generate a summarized video clip by the means of collecting shots.
In other words, a video skimming process is regarded as a video
reconstruction process. Starting from an empty output sequence, a
shot is recruited each time to the output sequence until the target
skimming ratio is achieved. The duration of the output video is
thus controlled by recruiting different amounts of video shots to
satisfy arbitrary skimming ratio. The recruiting order plays an
important role in the final result.
[0087] As discussed hereinabove, all the three aspects in video
skimming are considered: efficiency, saliency, and smoothness.
Given the requirements, we design several rules and propose a
"reconstruction reference tree" structure for our skimming
algorithm.
[0088] To maintain concept integrity, some embodiments require that
each concept contributes shots to the skimmed video. By having each
concept contribute shots to the skimmed video, the skimmed video
reflects the diversity of concepts of the original video and, thus
yields the maximum entropy, even though some concepts may seem not
salient. In some embodiments, it is possible to have less salient
shots added into the video skimming preview as a way of trading of
concept integrity and saliency maximization.
[0089] In embodiments, the concept importance is used as a factor
for deciding the recruiting order of different concept patterns. It
is not equivalent to the concept saliency. Rather, concept
importance is a more high-level argument that can reveal a video
producer's intention for the concepts' representation. Most
commonly, if the producer gives a long shot for a concept pattern,
or repeats the concept in many shots, then this concept can be
considered to be of high importance intentionally. Under this
assumption, the concept importance can be expressed as:
I.sub.l={.SIGMA.N.sub.k|Shot.sub.k C.sub.l},
[0090] where N.sub.k is the total number of frames in shot k within
concept 1. In an embodiment reconstruction framework, a shot is
first picked from the most important concept.
[0091] To increase or maximize saliency, in some embodiments, every
shot is assigned an average audio-visual saliency value to indicate
how exciting this shot is to viewers. Some shots, however, have
mismatched audio-visual concepts. For example, a video of two
people, A and B are talking; most shots will consistently show the
person's figure and play the person's voice. Some shots will show
A's figure while play B's voice. The case is rare but possible, and
we call it a mismatch. After the concept registration, the mismatch
can be easily found by comparing the registered spectral clustering
results.
[0092] When there is a mismatch, the audio-visual saliency of the
shot is decreased in some embodiments, since recruiting such a shot
may cause some misunderstanding to viewers. The audio-visual
saliency is reduced according to the following expression:
AvgSal.sub.k=.lamda.AvgSal.sub.k.sup.v+(1-.lamda.)AvgSal.sub.k.sup.a
AvgSal.sub.k=AvgSal.sub.k-.alpha.d.sub.k,
where .lamda. is weighing parameters to balance audio-visual
saliency and a is saliency penalty for audio-visual concepts
mismatch.
[0093] The most salient shot in each concept is defined as a
"must-in" shot, which means that, these shots are recruited in the
skimmed video in spite of the skimming ratio. This helps guarantee
concept integrity. The other shots are "optional" shots in the
sense that they can be recruited or not depending on the target
skimming ratio.
[0094] The reconstruction reference tree (RRT) is a tree structure
for video reconstruction guidance. According to embodiments, the
RRT is built according embodiments principles regarding concept
integrity, concept importance, shot saliency penalty for
audio-visual concept mismatch and "must-in" shots and optional
shots for each concept.
[0095] In an embodiment, the root of the RRT is the video concept
space, which is the learned through the spectral clustering
process. The first level leaves are the concepts, which are sorted
in importance descending order from left to right, and the second
level leaves are the shots. Under each concept, the shots are
sorted in saliency descending order from top to bottom.
[0096] The first child of each concept is the "must-in shot" and
the rest of the shots are optional shots. Since each concept may
have different number of shots, some "virtual shots" with zero
saliency are included to form an array of all shots. The resulting
array is called the shot table.
[0097] Turning to FIG. 8, embodiment RRT 400 is illustrated. Shot
table 402 has shots ordered according to concept 404 and shot
saliency. In an embodiment, concept categories 404 are ordered
according to concept importance. Similarly, shots within each
category are ordered according to saliency, such that the more
salient shots are ordered at the top of each concept category and
the least salient shots are ordered toward the bottom of each
concept category. In an embedment, ordered shots within each
category can include must-in shots 406, optional shots 408 and
virtual shots 410 as described above.
[0098] Given the RRT and shot table, the reconstruction process is
relatively straightforward. The following describes an embodiment
reconstruction algorithm:
TABLE-US-00001 ALGORITHM 1 (Reconstruction) Input: RRT, Target
skimming ratio R.sub.t Output: skimmed video V.sub.im, Current
skimming ratio R.sub.c Initialization: skimmed video = Empty;
Current skimming ratio R.sub.c = 0; Begin: Recruit a shot (must-in
or optional, skip the virtual shots) in the shot table in raster
scan order. Update R.sub.c; If R.sub.c.gtoreq.R.sub.t and all
must-in shots are recruited Stop; Else Loop; End
[0099] As the reconstruction is based on shots, the current
skimming ratio R.sub.c may not perfectly equal to the target
skimming ratio. In some embodiments, it may be more likely that
R.sub.c is slightly larger than R.sub.t (due to the stop criteria).
In order to precisely control the output video duration, pure
frame-level skimming, which is based on the attention model, is
used as post processing. The audio-visual saliency of a frame as is
computed as:
Sal.sub.t=.lamda.Sal.sub.t.sup.v+(1-.lamda.)Sal.sub.t.sup.a.
[0100] The audio-visual saliency of every frame that appears in the
output sequence is checked again. By thresholding on the saliency
curve, frames with relatively low saliency are discarded, thereby
allowing the final duration of the output video satisfy the target
duration. In addition, the smoothness requirement is also
considered to yield a viewer-friendly skimmed video. A
morphological smoothing operation is adopted which includes
deleting curve segments that are too short than K frames, and
joining together curve segments that are less than K frames apart.
In some embodiment, K is generally a small number, for example, 10
frames. Alternatively, other numbers can be used for K. The post
processing algorithm is described as follows:
TABLE-US-00002 ALGORITHM 2 (post processing) Input: V.sub.im,
R.sub.c, R.sub.t Output: final skimmed video V.sub.o
Initialization: saliency curve formation of V.sub.im using the user
attention model calculate curve preserving ratio R =
R.sub.t/R.sub.c Begin: smooth the saliency curve using median
filter calculate a threshold T such that R percent of curve are on
top of the threshold. thresholding V.sub.imusing T, do
morphological smoothing and the remaining frames compose V.sub.o
End.
[0101] FIG. 9 illustrates an embodiment saliency curve 450 with
threshold 452 to obtain a preserving ratio 95%. It should be
appreciated that in alternative embodiments, other thresholds can
be applied to obtain other threshold ratios.
[0102] The generation of embodiment video skimming previews
described hereinabove can be implemented in system 500 as shown in
FIG. 10. Referring to that figured, shot segmentor 502 is
configured to segment a video in to individual shots. The saliency
of each shot is determined by video saliency determination block
504, the output of which, is used by key frame determining block
506 to determine a key frame from among each shot. Visual feature
extractor block 508 is configured to extract visual features, and
visual word clustering block 509 is configured to cluster visual
words form visual concepts using methods described above. Shot
clustering block 510 is configured to cluster shots based on
different visual and audio descriptors to build concept
patterns.
[0103] In an embodiment, audio feature determination block 516 is
configured to determine audio features from each segmented shot,
and audio saliency determination block 518 is configured to the
saliency of each audio feature. Audio clustering is performed by
audio word clustering block 520 to produce audio concepts.
Furthermore, audio and visual concepts are aligned by block
522.
[0104] In an embodiment, reconstruction reference tree generation
block 512 creates an RRT based on saliency and concept importance,
according to embodiments described herein. Moreover, video skimming
preview generator 514 is configured to generate the skimming
preview.
[0105] FIG. 11 illustrates a processing system 600 that can be
utilized to implement methods of the present invention. In this
case, the main processing is performed in by processor 602, which
can be a microprocessor, digital signal processor or any other
appropriate processing device. Program code (e.g., the code
implementing the algorithms disclosed above) and data can be stored
in memory 604. The memory can be local memory such as DRAM or mass
storage such as a hard drive, optical drive or other storage (which
may be local or remote). While the memory is illustrated
functionally with a single block, it is understood that one or more
hardware blocks can be used to implement this function.
[0106] In one embodiment, the processor can be used to implement
various some or all of the units shown in FIG. 11. For example, the
processor can serve as a specific functional unit at different
times to implement the subtasks involved in performing the
techniques of the present invention. Alternatively, different
hardware blocks (e.g., the same as or different than the processor)
can be used to perform different functions. In other embodiments,
some subtasks are performed by the processor while others are
performed using a separate circuitry.
[0107] FIG. 11 also illustrates I/O port 606, which can be used to
provide the video to and from the processor. Video source 608, the
destination of which is not explicitly shown, is illustrated in
dashed lines to indicate that it is not necessary part of the
system. For example, the source can be linked to the system by a
network such as the Internet or by local interfaces (e.g., a USB or
LAN interface). In some embodiments, display device 612 is coupled
to I/O port 606 and supplies display 616, such as a CRT or
flat-panel display with a video signal, and audio device 610 is
coupled to I/O port 606 and drives acoustic transducer 614 with an
audio signal. Alternatively, audio device 610 and display device
612 can be coupled via a computer network, cable television
network, or other type of network. In a further embodiment, video
skimming previews can be generated offline and included in portable
media such as DVDs, flash drives, and other types of portable
media.
[0108] FIG. 12 provides a listing of notations used herein.
[0109] As discussed above, the present application provides a
number of new features including, but not limited to, using a
hierarchical video summarization framework for arbitrary and
accurate skimming ratio control with integrated concept
preservation. Embodiments also include the ability to provide
progressive video reconstruction from concept groups for high-level
summarization, concept group categorization by spectral clustering
of video shots, and alignment of audio concept groups with video
concept groups. In some embodiments, visual and audio Bag-of-Words
techniques are used for feature extraction, where visual words are
constructed by using a SIFT (Scale-invariant feature transform),
and audio words constructed by using Gabor-dictionary based
Matching Pursuit (MP) techniques.
[0110] In embodiments, saliency masking is used to provide for
robust and distinguishable Bag-of-Word feature extraction. In some
embodiments, visual saliency curve shaping uses a PQFT+dominant
motion contrast attention model, and audio saliency curve shaping
is performed using by low-level audio features, such as Maximum
Absolute Value, Spectrum Centroid, RMS, and ZCR. Furthermore,
saliency-curve based skimming is used as a low level
summarization.
[0111] An advantage of embodiments using spectral clustering is
that spectral clustering favors the classification of
locally-correlated data into one cluster because it adds another
constraint to distinguish the close-located or locally-connected
data and increase their similarity to be divided into one group. By
this constraint, the clustering result approaches human intuition
that a cluster with consistent members is generally subject to a
concentrated distribution. A further advantage of spectral
clustering is that the clustering result is not sensitive to the
number of members in the clusters.
[0112] While this invention has been described with reference to
illustrative embodiments, this description is not intended to be
construed in a limiting sense. Various modifications and
combinations of the illustrative embodiments, as well as other
embodiments of the invention, will be apparent to persons skilled
in the art upon reference to the description. It is therefore
intended that the appended claims encompass any such modifications
or embodiments.
* * * * *