U.S. patent application number 11/246607 was filed with the patent office on 2007-11-22 for texture replacement in video sequences and images.
This patent application is currently assigned to AT&T Corp.. Invention is credited to Adriana Dumitras, Barin Geoffry Haskell.
Application Number | 20070268301 11/246607 |
Document ID | / |
Family ID | 26930735 |
Filed Date | 2007-11-22 |
United States Patent
Application |
20070268301 |
Kind Code |
A1 |
Dumitras; Adriana ; et
al. |
November 22, 2007 |
TEXTURE REPLACEMENT IN VIDEO SEQUENCES AND IMAGES
Abstract
Systems and methods for reducing bit rates by replacing original
texture in a video sequence with synthesized texture. Reducing the
bit rate of the video sequence begins by identifying and removing
selected texture from frames in a video sequence. The removed
texture is analyzed to generate texture parameters. New texture is
synthesized using the texture, parameters in combination with a set
of constraints. Then, the newly synthesized texture is mapped back
into the frames of the video sequence from which the original
texture was removed. The resulting frames are then encoded. The bit
rate of the video sequence with the synthesized texture is less
than the bit rate of the video sequence with the original texture.
Also, the ability of a decoder to decode the new video sequence is
not compromised because no assumptions are made about the texture
synthesis capabilities of the decoder.
Inventors: |
Dumitras; Adriana;
(Sunnyvale, CA) ; Haskell; Barin Geoffry;
(Mountain View, CA) |
Correspondence
Address: |
AT&T CORP.
ROOM 2A207
ONE AT&T WAY
BEDMINSTER
NJ
07921
US
|
Assignee: |
AT&T Corp.
New York
NY
|
Family ID: |
26930735 |
Appl. No.: |
11/246607 |
Filed: |
October 7, 2005 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
11179701 |
Jul 12, 2005 |
|
|
|
11246607 |
Oct 7, 2005 |
|
|
|
10237489 |
Sep 9, 2002 |
6977659 |
|
|
11179701 |
Jul 12, 2005 |
|
|
|
60328627 |
Oct 11, 2001 |
|
|
|
Current U.S.
Class: |
345/582 ;
375/E7.081; 375/E7.083; 375/E7.085; 375/E7.182; 375/E7.185 |
Current CPC
Class: |
G06T 7/41 20170101; H04N
19/17 20141101; G06T 9/001 20130101; G06T 11/001 20130101; H04N
19/27 20141101; G06T 7/40 20130101; H04N 19/14 20141101; H04N
19/186 20141101; H04N 19/167 20141101; G06T 9/00 20130101; H04N
19/23 20141101 |
Class at
Publication: |
345/582 |
International
Class: |
G09G 5/00 20060101
G09G005/00 |
Claims
1.-7. (canceled)
8. A method for processing video data, the method comprising:
removing texture from at least some frames of a set of frames in a
video sequence; analyzing the removed texture to obtain texture
parameters; synthesizing new texture based on the texture
parameters and at least one of the following qualitative
constraints: coefficient magnitude correlations, cross scale
statistics, coefficient correlations, marginal statistics, dominant
texture orientation, overall color, and color saturation related to
whether an original texture associated with the set of frames is
structure and busy or unstructured and smooth, wherein the original
texture is structured and busy, then synthesizing the new texture
is based on the following constraints: marginal statistics,
coefficient correlations, coefficient magnitude correlations,
cross-scale statistics, overall color and color saturation; and
wherein if the original texture is instructed and smooth, then
synthesizing the new texture is based on at least the constraints:
coefficient magnitude correlations, cross-scale statistics, overall
color and color saturation; and inserting the synthesized new
texture into the set of frames to enable display of the set of
frames.
9. The method of claim 8, wherein the synthesized new texture is
visually similar to the original texture.
10. (canceled)
11. (canceled)
12. The method of claim 8, wherein the texture is synthesized
according to whether the original texture is structured and busy or
unstructured and smooth.
13.-14. (canceled)
15. The method of claim 8, wherein the synthesized new texture is
similar to but distinguishable from the removed texture.
16. The method of claim 8, wherein a bit rate of a video sequence
that includes the synthesized texture is less than a bit rate of an
original video sequence.
17. The method of claim 8, wherein a region of interest is
identified within texture of an initial frame.
18. The method of claim 17, wherein a mean vector is computed for
the region of interest.
19. The method of claim 18, wherein the texture is synthesized and
inserted further according to a method comprising: computing an
angular map and a modulus map for an initial frame, wherein the
angular map and the modulus map represent color characteristics;
comparing color characteristics of the region of interest with
color characteristics of all pixels in the initial frame; and
including in the texture to be removed pixels whose color
characteristics are similar to the color characteristics of the
region of interest.
20.-31. (canceled)
32. A tangible computer-readable medium storing instructions for
controlling a computing device to process video data, the
instructions comprising: removing texture from at least some frames
of a set of frames in a video sequence; analyzing the removed
texture to obtain texture parameters; synthesizing new texture
based on the obtained texture parameters and at least one of the
following qualitative constraints: coefficient magnitude
correlations, cross scale statistics, coefficient correlations,
marginal statistics, dominant texture orientation, overall color,
and color saturation related to whether an original texture
associated with the set of frames is structure and busy or
unstructured and smooth, wherein the original texture is structured
and busy, then synthesizing the new texture is based on the
following constraints: marginal statistics, coefficient
correlations, coefficient magnitude correlations, cross-scale
statistics, overall color and color saturation; and wherein if the
original texture is instructed and smooth, then synthesizing the
new texture is based on at least the constraints: coefficient
magnitude correlations, cross-scale statistics, overall color and
color saturation; and inserting the synthesized new texture into
the set of frames.
33. The computer-readable medium of claim 32, wherein synthesized
new texture is visually similar to the original texture.
34.-38. (canceled)
39. The computer-readable medium of claim 32, wherein synthesized
new texture is similar to but distinguishable from the removed
texture.
40. The computer-readable medium of claim 32, wherein a bit rate of
a video sequence that includes the synthesized texture is less than
a bit rate of an original video sequence.
41. The computer-readable medium of claim 32, wherein a region of
interest is identified within texture of an initial frame.
42. The computer-readable medium of claim 41, wherein a mean vector
is computed for the region of interest.
43. The computer-readable medium of claim 42, wherein the texture
is synthesized and inserted further according to a method
comprising: computing an angular map and a map for an initial
frame, wherein the angular map and the modulus map represent color
characteristics; comparing color characteristics of the region of
interest with color characteristics of all pixels in the initial
frame; and including in the texture to be removed pixels whose
color characteristics are similar to the color characteristics of
the region of interest.
Description
[0001] This application claims the benefit of U.S. Provisional
Patent Application Ser. No. 60/328,627, entitled "A Texture
Replacement Method at the Encoder for Bit Rate Reduction of
Compressed Video," filed Oct. 11, 2001, which is incorporated
herein by reference.
RELATED APPLICATIONS
[0002] This application is related to commonly assigned U.S. patent
application Ser. No. ______ entitled "System And Method For
Encoding And Decoding Using Texture Replacement," filed ______ by
Adriana Dumitras and Barin Geoffry Haskell and claiming priority to
U.S. Provisional Patent Application Ser. No. 60/360,027, filed Feb.
12, 2002. This above-identified application is incorporated by
reference herewith.
BACKGROUND OF THE INVENTION
[0003] 1. The Field of the Invention
[0004] The present invention relates to systems and methods for
reducing a bit rate of a video sequence. More particularly, the
present invention relates to systems and methods for reducing a bit
rate of a video sequence by replacing original texture of the video
sequence with synthesized texture at the encoder.
[0005] 2. Background and Relevant Art
[0006] One of the goals of transmitting video sequences over
computer networks is to have a relatively low bit rate while still
maintaining a high quality video at the decoder. As technology
improves and becomes more accessible, more users are leaving the
realm of 56K modems and moving to Digital Subscriber Lines (DSL),
including VDSL and ADSL, which support a higher bit rate than 56K
modems. VDSL, for example, supports bit rates up to 28
Mbits/second, but the transmission distance is limited. The maximum
transmission distance for a 13 Mbits/second bit rate is 1.5 km
using VDSL. ADSL, on the other hand, can support longer distances
using existing loops while providing a bit rate of approximately
500 kbits/second.
[0007] Video standards, such as MPEG-2, MPEG-4, and ITU H.263, can
achieve bit rates of 3 to 9 Mbits/second, 64 kbits to 38.4
Mbits/second, and 8 kbits to 1.5 Mbits/second, respectively. Even
though video sequences with bit rates of hundreds of kbits/second
can be achieved using these standards, the visual quality of these
video sequences is unacceptably low, especially when the content of
the video sequences is complex.
[0008] Solutions to this problem use model-based analysis-synthesis
compression methods. Model-based analysis-synthesis compression
methods perform both analysis and synthesis at the encoder to
modify parameters in order to minimize the error between the
synthesized model and the original. The resulting parameters are
transmitted to the decoder, which is required to synthesize the
model again for the purpose of reconstructing the video
sequence.
[0009] Much of the model-based analysis-synthesis compression
methods have focused on modeling human head-and-shoulders objects
while fewer attempts have modeled background objects. Focusing on
human head-and-shoulder objects often occurs because in many
applications, such as videoconferencing applications, the
background is very simple. However, background modeling may also
achieve a significant reduction of the bit rate as the bit rate of
I (intra) frames is often dependent on the texture content of each
picture. To a lesser extent, the bit rate of B (bi-directionally
predicted) frames and P (predicted) frames is also affected by
texture content as moving objects uncover additional background
objects.
[0010] One proposal for reducing the bit rate is to use sprite
methods on the background objects. Sprites are panoramic pictures
constructed using all of the background pixels that are visible
over a set of video frames. Instead of coding each frame, the
sprite is compressed and transmitted. The background image can be
reconstructed using the sprite and associated camera motion
parameters. Sprite methods require exact object segmentation at the
encoder, which is often a difficult task for complex video
sequences. In addition, the motion or shape parameters that are
transmitted with the sprite consume some of the available bit rate.
These limitations may be addressed by filtering the textured areas.
Unfortunately, different filters must be designed for various
textures.
[0011] Texture replacement has also been proposed as a method of
background modeling. In one example, the original texture is
replaced with another texture that is selected from a set of
textures. However, this requires that the set of replacement
textures be stored at the encoder. In another example, the texture
of selected regions is replaced at the encoder with pixel values
that represent an "illegal" color in the YUV color space. At the
decoder, the processed regions are recovered using chroma keying.
There is an explicit assumption that texture synthesis, using
texture parameters sent from the encoder, followed by mapping of
the synthesized texture onto the decoded video sequences, is
performed at the decoder. This method therefore assumes that the
reconstruction is performed at the decoder using a method that is
dependent on the decoder's processing capabilities. The drawbacks
of these approaches are that the processing capabilities of the
decoder are assumed and that the computational costs of the
decoding stage are increased.
BRIEF SUMMARY OF THE INVENTION
[0012] These and other limitations of the prior art are overcome by
the present invention which relates to systems and methods for
reducing the bit rate of a video sequence through texture
replacement at the encoder. The capabilities of the decoder are not
assumed and the decoder is not required to perform texture analysis
and synthesis. The synthesized texture that replaces the original
texture has similar perceptual characteristics to the original
texture. Thus, the video sequence with the synthesized texture is
visually similar to the original video sequence. The bit rate of
the video sequence with synthesized textures is reduced because the
synthesized textures that have replaced the original textures can
be coded more effectively.
[0013] Texture replacement, in accordance with the present
invention, occurs at the encoder and is therefore independent of
the capabilities of the decoder. Texture replacement begins by
selecting and removing texture from some or all of the original
frames in the video sequence. The removed texture is analyzed to
obtain texture parameters. Then, new texture is synthesized using
the texture parameters in combination with a set of qualitative
constraints. The synthesized texture can be compressed more
effectively than the original texture and is also similar to, yet
distinguishable from, the original texture. After the new texture
is synthesized, the synthesized texture is inserted back into the
original frames and the video sequence that includes the
synthesized texture is encoded. Advantageously, the bit rate of the
compressed video sequence with the synthesized texture is lower
than the bit rate of the compressed video sequence with the
original texture.
[0014] Additional features and advantages of the invention will be
set forth in the description which follows, and in part will be
obvious from the description, or may be learned by the practice of
the invention. The features and advantages of the invention may be
realized and obtained by means of the instruments and combinations
particularly pointed out in the appended claims. These and other
features of the present invention will become more fully apparent
from the following description and appended claims, or may be
learned by the practice of the invention as set forth
hereinafter.
BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to describe the manner in which the above-recited
and other advantages and features of the invention can be obtained,
a more particular description of the invention briefly described
above will be rendered by reference to specific embodiments thereof
which are illustrated in the appended drawings. Understanding that
these drawings depict only typical embodiments of the invention and
are not therefore to be considered to be limiting of its scope, the
invention will be described and explained with additional
specificity and detail through the use of the accompanying drawings
in which:
[0016] FIG. 1 is a flow diagram illustrating an exemplary method
for reducing a bit rate of a video sequence by replacing original
texture with synthesized texture;
[0017] FIG. 2 illustrates a frame that includes various candidate
textures which can be replaced with synthesized textures;
[0018] FIG. 3 is a flow diagram of a method for removing original
texture from various frames of a video sequence; and
[0019] FIG. 4 illustrates a recursive transform used in image
decomposition to obtain texture parameters.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0020] The present invention relates to systems and methods for
texture replacement in video sequences and to reducing bit rates of
video sequences through texture replacement. The bit rate of a
compressed video sequence is reduced because the synthesized
texture can be more effectively compressed than the original
texture. Because the texture replacement occurs as part of the
encoding process, the computational costs of the decoder are not
increased and the processing/synthesizing capabilities of the
decoder are not assumed.
[0021] FIG. 1 is a flow diagram that illustrates one embodiment of
a method for replacing texture in a video sequence and for reducing
the bit rate of the encoded video sequence. Reducing the bit rate
of a video sequence begins by removing selected texture from a
frame or from a series of frames included in the video sequence
(102). Typically, texture from a single frame is identified, while
the identified texture from that particular frame is removed from a
set of frames. This eliminates the need to identify texture in each
frame, which would increase the computational costs of texture
replacement. After the original texture has been removed from the
selected frame, the removed texture is analyzed (104) to obtain
texture parameters. The texture parameters, in combination with a
set of constraints, are used to generate synthesized texture (106).
The synthesized texture is then inserted into the original frames
(108) from which the original texture was removed.
[0022] One advantage of texture replacement is that the synthesized
texture can be compressed more effectively than the original
texture. The ability to more effectively compress the synthesized
texture results in a reduced bit rate for the encoded video
sequence. The constraints are applied during the synthesis of the
new texture to ensure that the synthesized texture is similar to
the original texture. This is useful, for example, in order to
retain the artistic representation of the original frames. However,
the present invention does not require that the synthesized texture
resemble the original texture.
[0023] FIGS. 2 and 3 more fully illustrate how texture is selected
and removed from the original frames of a video sequence. FIG. 2
illustrates an exemplary frame 200. The frame 200 is a picture
that, in this example, includes a person 202, a person 203 and a
background. The frame 200 includes several textures that can be
selected for removal. Texture refers to those portions or regions
of a frame that have the same or similar characteristics and/or
appearance. Usually, the background is the primary source of
texture, although the present invention is not limited to
background textures. One reason for selecting background textures
is that these same textures or regions are present in multiple
frames and in substantially the same locations. For example, if no
movement is present in a particular set of frames, the background
will remain constant across the set of frames. However, the systems
and methods described herein can be applied even when a set of
frames includes moving objects.
[0024] In this example, the background includes several candidate
textures that can be selected for removal. The background of the
frame 200 includes the ground texture 204, the sky texture 206 and
the cloud texture 208. Each of these objects in the background are
examples of textures. It is likely that the region of the frame 200
covered by the ground texture 204, the sky texture 206, or the
cloud texture 208, respectively, have similar appearances or
characteristics.
[0025] When a region of texture is selected for removal, it is
useful to select a texture that has a spatial region of support
that covers a reasonable part of the frame area over a sufficiently
large number of frames. In one embodiment, a texture that covers an
area that is larger than 30% of the frame area is typically
selected, although the present invention can apply to any texture,
regardless of the frame area occupied by the selected texture. In
addition, selecting a texture that can be replaced in a large
number of frames is advantageous because it has an impact on the
bit rate of the encoded video sequence.
[0026] It is also useful to select a texture that belongs to a
class of textures that are amenable to replacement. Generally,
candidate textures include regions of a frame that have similar
characteristics and/or appearance. Exemplary textures include
natural textures (foliage, grass, ground, water, sky, building
facades, etc.), and the like that can be identified in video
sequences or test sequences such as movie sequences. Frames with
replaceable texture typically have absent or slow global motion
and/or only a few moving objects. Thus, the present invention can
be applied to particular frame sequences within a video sequence
and does not have to be applied to the entire video sequence.
Texture Removal and Region Segmentation
[0027] FIG. 3 illustrates a flow diagram illustrating how a texture
is selected and removed from the frame illustrated in FIG. 2.
First, a region of interest (ROI) 210 is selected within the
texture to be replaced (302). In this example, the selected ROI 210
is within the ground texture 204. The ROI 210 typically includes
more than one pixel. The ROI 210, for example, may include a
7.times.7 array of pixels. After the ROI 210 is selected, the color
characteristics of the ROI 210 are compared to the color
characteristics (304) of all pixels within the frame 200. The
pixels that have similar characteristics are classified or
identified as belonging to the same region (306) as the ROI 210.
The pixels identified in this manner are thus included in the
texture that will be removed from the frame(s). The pixels that are
in the identified region are removed and are temporarily replaced
with pixels that have an arbitrary value. In one embodiment, the
replacement pixels have a constant value equal to the mean color of
the selected and identified region.
[0028] More particularly in one embodiment, region segmentation and
texture removal occurs in stages. For example, let a color frame be
represented by a set of two-dimensional planes in the YUV color
space. Each of these image planes is represented as a matrix and
each matrix element includes a pixel value in row i and column j.
More specifically, the Y frames consist of all of the pixels {(i,
j), with 1.ltoreq.i.ltoreq.M, 1.ltoreq.j.ltoreq.N}. First, original
frames from the YUV color space are converted to the RGB color
space. Second, the location {i.sub.r, j.sub.r} is selected from a
region-of-interest, such as the ground texture 204 shown in FIG. 2,
and the ROI is constructed from: ROI r = { i = i r + k i , j = j r
+ k j } , with - [ w r 2 ] .ltoreq. k i , k j .ltoreq. [ w r 2 ] ,
( 1 ) ##EQU1## where the operator [ ] denotes "the integer part of"
and w.sub.r is odd. Alternatively, the region-of-interest of a size
equal to w.sub.r.times.w.sub.r pixels can be manually selected from
a region of interest. The pixel values are smoothed by applying an
averaging filter to the ROI 210 in each of the R, G and B color
planes, and the mean vector [.mu..sub.r.sup.R .mu..sub.r.sup.G
.mu..sub.r.sup.B] is computed, where .mu..sub.r.sup.R,
.mu..sub.r.sup.G, and .mu..sub.r.sup.B stand for the mean values
within the ROI in the R, G, and B color planes, respectively.
[0029] Next, the angular map and the modulus map of the frame are
computed as follows: .theta. .function. ( i , j ) = 1 - 2 .pi.
.times. arccos .function. ( v ( i , j ) .times. v ref v ( i , j )
.times. v ref ) , and ( 2 ) .eta. .function. ( i , j ) = 1 - v ( i
, j ) - v ref 3 255 2 , ( 3 ) ##EQU2## respectively, with respect
to a reference color vector. The following works more fully
describe angular and modulus maps and are hereby incorporated by
reference: (1) Dimitris Androutsos, Kostas Plataniotis, and
Anastasios N. Venetsanopoulos, "A novel vector-based approach to
color image retrieval using a vector angular-based distance
measure," Computer Vision and Image Understanding, vol. 75, no.
1/2, pp. 46-58, July/August 1999; (2) Adriana Dumitras and
Anastasios N. Venetsanopoulos, "Angular map-driven snakes with
application to object shape description in color images," accepted
for publication in IEEE Transactions on Image Processing., 2001;
and (3) Adriana Dumitras and Anastasios N. Venetsanopoulos, "Color
image-based angular map-driven snakes," in Proceedings of IEEE
International Conference on Image Processing, Thessaloniki, Greece,
October, 2001. Notations v(.sub.i,j) and v.sub.ref stand for a
color vector [R(i, j) G(i,j) B(i,j)] in the RGB color space, and
the reference vector that is selected to be equal to
[.mu..sub.r.sup.R .mu..sub.r.sup.G .mu..sub.r.sup.B] respectively.
The notation .theta. stands for the value of the angle given by (2)
between the vector v(.sub.i,j) and the reference vector v.sub.ref.
The notation .eta. stands for the value of the modulus difference
given by (3) between the vector v(.sub.i,j) and the reference
vector v.sub.ref, respectively.
[0030] In order to identify the pixels that have similar color
characteristics to those of the reference vector, the distance
measure is computed by d.sup..theta..eta.(i,j)=exp
[-.theta.(i,j).eta.(i,j)] (4) and the mean distance given is
computed by .mu. .times. d r = .times. { d .theta..eta. .function.
( i , j ) } , ( 5 ) ##EQU3## where notation .epsilon. stands for a
mean operator over ROI.sub.r. All of the pixels within the frame
that satisfy the constraint ( d .theta..eta. .function. ( i , j ) -
.mu. .times. d r ) 2 .times. c ( 6 ) ##EQU4## are clustered into
regions.
[0031] Next, the segmented regions that satisfy the constraint of
(6) are identified and labeled. In one embodiment, all regions with
a normalized area A.sub.R smaller than a threshold area, i.e., Ar
MN .ltoreq. .times. .times. A ##EQU5## where M.times.N is the frame
area, are discarded. The remaining regions are labeled. If a
segmentation map consisting of all of the segmented regions is not
considered acceptable by the user, another ROI location is selected
and processed as described above. The labeled regions are removed
from the frame and texture removal of the current frame is
complete. The frames having the texture and color removed using the
segmentation map are obtained at the end of the texture removal
stage.
[0032] To summarize, the region segmentation and texture removal
procedures described above identify pixels in the original frame
that have similar characteristics in terms of color to those of the
pixel (location) or ROI selected by the user. The color
characteristics of the identified pixels or segmented regions are
evaluated using an angular map and a modulus map of the color
vectors in the RGB color space. Angular maps identify significant
color changes within a picture or frame, which typically correspond
to object boundaries, using snake models. To identify color changes
or boundaries, simple gray-level edge detection is applied to the
angular maps. Identification of major change of direction in vector
data and color variation computation by differences between inside
and outside contour points can also be applied. In the present
invention, however, a classification of the pixels is performed in
each frame using the angular map. Despite the fact that the
performance of the color-based region segmentation stage depends on
the selection of the threshold values .epsilon.C, .epsilon.A and
the size of the ROI, the values of these parameters may be
maintained constant for various video sequences.
Texture Analysis
[0033] Analyzing the removed texture includes computing a set of
texture parameters from the removed texture. In one embodiment, a
parametric statistical model is employed. This model, which employs
an overcomplete multiscale wavelet representation, makes use of
steerable pyramids for image decomposition. Steerable pyramids as
known in the art are more fully discussed in the following article,
which is hereby incorporated by reference: Javier Portilla and Eero
P. Simoncelli, "A parametric texture model based on joint
statistics of complex wavelet coefficients," International Journal
of Computer Vision, vol. 40, no. 1, pp. 49-71, 2000. The
statistical texture descriptors are based on pairs of wavelet
coefficients at adjacent spatial locations, orientations and scales
(in particular, the expected product of the raw coefficient pairs
and the expected product of their magnitudes), pairs of
coefficients at adjacent scales (the expected product of the fine
scale coefficient at adjacent scale coefficient), marginal
statistics and lowpass coefficients at different scales.
[0034] First, a steerable pyramid decomposition is obtained by
recursively decomposing the texture image into a set of oriented
subbands and lowpass residual band. The block diagram of the
transform is illustrated in FIG. 4, where the area enclosed by the
dashed box 400 is inserted recursively at the point 402. Initially,
the input image is decomposed into highpass and lowpass bands using
the exemplary filters L 0 .function. ( r , .theta. ) = L .function.
( r 2 , .theta. ) 2 .times. .times. and .times. .times. H 0
.function. ( r , .theta. ) = H .function. ( r 2 , .theta. ) .
##EQU6## The lowpass band is then decomposed into a lower frequency
band and a set of oriented bands. The filters used in this
transformation are polar-separable in the Fourier domain and are
given by: L .function. ( r , .theta. ) = { 2 .times. cos .function.
( .pi. 2 .times. log 2 .function. ( 4 .times. r .pi. ) ) , .pi. 4
.times. r .times. .pi. 2 2 , r .ltoreq. .pi. 4 , and 0 , r .gtoreq.
.pi. 2 .times. .times. B k .function. ( r , .theta. ) = H
.function. ( r ) .times. G k .function. ( .theta. ) , k .times.
.times. .times. 0 .xi. .times. K - 1 ##EQU7## where H .function. (
r ) = { cos .function. ( .pi. 2 .times. log 2 .function. ( 2
.times. r .pi. ) ) , .pi. 4 .times. r .times. .pi. 2 1 , r .gtoreq.
.pi. 2 , and 0 , r .ltoreq. .pi. 4 .times. .times. G k .function. (
.theta. ) = { 2 K - 1 .times. ( K - 1 ) ! K .function. [ 2 .times.
( K - 1 ) ] ! .function. [ cos .function. ( .theta. - .pi. .times.
.times. K K ) ] K - 1 , .theta. - .pi..kappa. K .times. .pi. 2 0 ,
otherwise . ##EQU8##
[0035] The notations r and .theta. stand for polar coordinates in
the frequency domain, K denotes the total number of orientation
bands.
[0036] Statistical texture descriptors are computed using the image
decomposition previously obtained. More specifically, marginal
statistics, correlations of the coefficients, correlations of the
coefficients' magnitudes and cross-scale phase's statistics are
computed. In terms of marginal statistics (a) the skewness and
kurtosis of the partially reconstructed lowpass images at each
scale, (b) the variance of the highpass band, (c) the mean
variance, skewness and kurtosis and (d) the minimum and maximum
values of the image pixels (range) are computed at each level of
the pyramid. In terms of coefficient correlations, the
autocorrelation of the lowpass images computed at each level of the
pyramid decomposition is computed. In terms of magnitude
correlation, the correlation of the complex magnitude of pairs of
coefficients at adjacent positions, orientations and scales is
computed. More specifically, (e) the autocorrelation of magnitude
of each subband, (f) the crosscorrelation of each subband
magnitudes with those of other orientations at the same scale, and
(g) the crosscorrelation of subband magnitudes with all
orientations at a coarser scale are obtained. Finally, in terms of
cross-scale statistics, the complex phase of the coarse-scale
coefficients is doubled at all orientations and then the
crosscorrelation between these coefficients and the fine-scale
coefficients is computed. Doubling of the coarse-scale coefficients
is motivated by the fact that the local phase of the responses to
local feature such as edges or lines changes at a rate that is, for
fine-scale coefficients, twice the rate of that of the coefficients
at a coarser scale.
Texture Synthesis
[0037] During texture synthesis, a synthesized texture is created
that can be compressed more effectively than the original texture.
In one embodiment, the synthesized texture is similar to, yet
distinguishable from, the original texture. Because the synthesized
texture can be compressed more effectively, the bit rate of the
compressed frames with synthesized texture is lower than the bit
rate of the compressed frames with the original texture. Retaining
the visual similarity ensures that the frames with synthesized
texture convey the same artistic message as that of the frames with
the original texture.
[0038] A set of qualitative texture synthesis constraints are used
to achieve visual similarity between the synthesized texture and
the original texture. The texture parameters selected for texture
synthesis are derived using the set of qualitative constraints.
[0039] In one embodiment, the synthesis of a new texture that is
visually similar to the original texture is subject to constraints
in terms of dominant texture orientation, and overall color and
color saturation. Exemplary constraints include, but are not
limited to:
[0040] (C1) Marginal statistics;
[0041] (C2) Coefficient correlations;
[0042] (C3) Coefficient magnitude correlations;
[0043] (C4) Cross-scale statistics;
[0044] (C5) Overall color; and
[0045] (C6) Color saturation.
[0046] If the original texture that has been removed from the
original frames is structured and busy, the synthesized texture is
subject to the constraints C1, C2, C3, C4, C5 and C6. If the
original texture is unstructured and smooth, the synthesized
texture is subject to the constraints C3, C4, C5 and C6.
[0047] Using these constraints, the new texture is synthesized by
first decomposing an image containing Gaussian white noise using a
complex steerable pyramid. Next, a recursive coarse-to-fine
procedure imposes the statistical constraints on the lowpass and
bandpass bands while simultaneously reconstructing a lowpass image.
In one embodiment of the present invention, the order in which the
constraints are applied is C3, C4, C2, C1, C5, and C6 for
structured and busy textures, and C3, C4, C5, and C6 for
unstructured and smooth textures.
[0048] The texture synthesis constraints are derived from the basic
categories (vocabulary) and rules (grammar) used by humans when
judging the similarity of color patterns. The following work, which
more fully describes the vocabulary and grammar of color patterns
is hereby incorporated by reference: Aleksandra Mojsilovic, Jelena
Kovacevic, Jianying Hu, Robert J. Safranek, and S. Kicha Ganapathy,
"Matching and retrieval based on the vocabulary and grammar of
color patterns," IEEE Transactions on Image Processing, vol. 9, no.
1, pp. 38-54, January 2000. In one embodiment, the basic pattern
categories used by humans for similarity evaluation of color
patterns are directionality, regularity and placement, complexity
and heaviness and the basic color categories used by humans for the
same task are overall color and color purity. The following table
provides the meanings for these criteria: TABLE-US-00001 Derived
Criterion Similarity for constrained Criterion Expresses texture
synthesis Overall Perception of a single dominant color, Preserve
the color or a multicolored image that creates overall color in the
impression of a dominant color the synthesized texture Direction-
Perception of the dominant orientation Preserve the ality and in
edge distribution, or the dominant dominant orientation direction
in the repetition of the orientation in structural element the
synthesized texture Regularity perception of the placement (and
small Not applied and perturbation in placement), repetition
placement and uniformity of the patterns Color purity Perception of
the degree of Preserve the color (overall colorfulness in patterns
saturation in the saturation) synthesized texture Pattern
Perception of a general impression Preserve the type complexity
based on the type of overall color of the overall and (light versus
dark) or the overall color (light or heaviness chroma (saturated
versus unsaturated) dark) or the spatial frequency in the
repetition of the structural element or the color contrast
[0049] Because different visual pathways process patterns and
colors in the human visual system, separate constraints for texture
and color are derived in the texture synthesis process. Moreover,
texture is typically synthesized with gray levels, i.e. luminance
frames, and color constraints are imposed on the chrominance
frames. In terms of texture, the synthesized texture should have
similar dominant directionality as that of the original texture. In
terms of color, the synthesized texture should have similar overall
color and color saturation as those of the original texture.
[0050] To address the texture requirement, the magnitude
correlation constraint (C3) which represents the structure in
images and the cross-scale statistics constraint (C4) which allows
distinguishing lines and edges are selected and applied. While
these texture constraints are sufficient for unstructured and
smooth textures, they do not allow synthesizing structured and busy
textures with appearances that are similar to those of the original
ones. Therefore, for structured and busy textures, the marginal
statistics constraint (C1) and the coefficient correlation
constraint (C2) are used. The correlation constraint characterizes
the regularity of the texture as represented by periodic or
globally oriented structure in the set of constraints. The overall
color (C5) and color saturation (C6) are the color constraints that
address the color requirements. The overall color constraint may be
further expressed in terms of similar color moments (mean color and
standard color deviation). Because the pixel values in the
segmented regions have been replaced with mean pixel values within
the region, the overall color is easily preserved. Because
preserving only the mean color would yield textures with discolored
appearances in the cases of textures that exhibit large color
variations, the color situations are also preserved.
[0051] By selecting different sets of constraints for unstructured
and smooth textures, and structured and busy textures, different
characteristics of the background textures that are present in
video sequences can be utilized. For example, by removing the
marginal statistics constraint (C1), synthesized textures that
differ in their first and second order statistics from the original
textures can be obtained. Consequently, according to Julesz's
conjecture, the synthesized and original textures would then be
distinguishable in pre-attentive (undetailed) evaluation of less
than 50 milliseconds (ms). At a frame rate of 24 frames per second,
each frame is indeed viewed in a pre-attentive mode, i.e., for
approximately 40 (<50) ms. By removing the coefficient
correlation constraint (C2) for the same class of unstructured and
smooth textures, the requirement that the synthesized and the
original textures have similar periodicity is relaxed. This, in
turn, improves the compression effectiveness.
[0052] The texture synthesized is mapped on the luminance frame of
the video sequence. The segmented regions within the chrominance
frames remain filled with pixels having mean values as obtained by
the region segmentation. Thus, the color constraints are
satisfied.
[0053] Although the present invention has been described in terms
of a video sequence, it is understood by one of skill in the art
that the systems and methods described herein also apply to a still
image or to a single frame. Thus removing texture from a set of
frames includes the case where the set of frames is a single frame
or a still image.
[0054] The present invention may be embodied in other specific
forms without departing from its spirit or essential
characteristics. The described embodiments are to be considered in
all respects only as illustrative and not restrictive. The scope of
the invention is, therefore, indicated by the appended claims
rather than by the foregoing description. All changes which come
within the meaning and range of equivalency of the claims are to be
embraced within their scope.
* * * * *