Image analysis – Color image processing – Pattern recognition or classification using color
Reexamination Certificate
1999-09-24
2003-03-18
Tran, Phuoc (Department: 2621)
Image analysis
Color image processing
Pattern recognition or classification using color
C382S164000, C382S167000, C382S173000, C382S180000, C382S224000, C358S515000, C358S518000
Reexamination Certificate
active
06535633
ABSTRACT:
BACKGROUND OF THE INVENTION
The present invention is directed to the art of digital image processing and, more particularly, to a method and apparatus for overriding segmentation classification information generated during single channel segmentation of a multi-channel color image. The present invention is especially well suited for addressing the white/black detection failure modes prevalent in single channel segmentation schemes by selectively re-classifying appropriate pixels as an “other” class and will be described with particular reference thereto. However, it is to be understood that the present invention has broader application and can be used to compensate for a wide range of segmentation failure modes and can be used with a wide variety of digital images and other digital information or data.
Segmentation is an important area of electronic image processing, particularly color image processing. In systems that use image segmentation information such as page description language (PDL) or other document description type systems, a plurality of predefined classes detected in compound documents are used for image storage, processing, rendering, and for other functions. The set of predefined classes includes text, halftone, contone, line art, picture, and many more. It is important, therefore, for the segmentation scheme used in classifying the color image to be able to correctly recognize and accurately identify pixels belonging to the set of classification groups supported by the PDL. That is, it is important for the segmentation technique to be matched to the document description system so that each pixel is classified as one of a plurality of predetermined recognized classification groups.
Sometimes, one or more pixels in a color image cannot be classified as belonging to one of the plurality of predetermined classification groups with a reasonable certainty. In that case, some prior art segmentation schemes label the indeterminate pixel as belonging to an “other” class. As an example, if a segmentation scheme can classify color image pixels as belonging to either a text or contone group only, and then halftone pixels are encountered in the digital color image, the segmentation processing would be unable to label the halftone pixels as either text or contone with a reasonable degree of certainty. One solution is to incorrectly label the halftone pixels as being either a text or contone type class even though the halftone pixel belongs to neither. Another solution is to merely label the halftone pixel as belonging to an “other” class whereupon similar pixels of indeterminate class must be further processed or handled in some unique manner. The use of the “other” class in color image segmentation schemes for identifying portions of an image having indeterminate classification characteristics has proven to be effective.
In more complicated segmentation schemes
10
, such as the one shown at
FIG. 1
, each color space of a color input image
12
is processed by a separate independent full scale single channel segmentation processing unit that operates on each pixel of the multi-channel color input image simultaneously. In the example shown in
FIG. 1
, each pixel in the red color space R is processed by a first segmentation processing unit
14
. Similarly, each pixel in the green and blue color spaces G, B is processed by an independent segmentation processing unit
16
,
18
, respectively. Each of the segmentation processing units generates a pixel classification output signal for use by a discriminator circuit
20
to appropriately label the processed pixel as belonging to one of a plurality of predetermined segmentation classes. In the segmentation scheme
10
shown in the figure, each segmentation processing unit operates simultaneously and in parallel to process individual pixels of the color input image
12
one at a time until each of the pixels are provided with an appropriate segmentation class tag for use downstream in the PDL.
Although the segmentation scheme
10
shown in
FIG. 1
is fairly robust, it is quite complicated and somewhat costly to implement. Accordingly, a technique known as “single channel segmentation” for classifying pixels of color input images has been developed.
FIG. 2
shows an example of a prior art single channel color image segmentation circuit
22
. As illustrated there, only a single segmentation processing unit
24
is used to generate pixel classification signals for use by the discriminator circuit
26
to generate the appropriate pixel class tags. Typically, a single channel for input into the segmentation processing unit
24
is created from the multiple color channels, e.g. R, G, B of the color input image
28
using a technique known as projection. In projection, an inner product is determined between the multi-channel input video and a single predetermined direction. As an example, if the input video is an RGB image, the video value of each pixel can be represented by V
in
=[R
in
G
in
B
in
]′. A single channel for segmentation of the video, S
V
, can be determined from S
V
=W′*V
in
where W is a weighting vector: W=[W
1
W
2
W
3
]′. In order to ensure that the output is limited to 8 bits, the weighing vector is typically normalized by &Sgr;
i
w
i
=1.
One of the problems associated with single channel detection schemes is that occasionally very important information is lost during the conversion from three to one channel. As noted above, however, it is costly to include all three channels in performing segmentation. Accordingly, the single channel segmentation scheme shown in
FIG. 2
is often employed where cost is a concern and where segmentation accuracy is not a major concern.
Typical prior art single channel segmentation schemes are prone to several failure modes including classifying certain bright colors (e.g. yellow) as the class “white” and certain dark colors (e.g. purple) as the class “black”. White and black detection are the predominant failure modes. Typically, white is used as background in color images. Black is used as a key identifier in text recognition and to make decisions in certain compression schemes. In image processing, the white/color background class is very important. The black/white background class is used often to separate various regions within the color input image, for compression algorithms where background information (white/black class) is heavily compressed to separate windows within the color input image wherein the various windows are stored and/or processed independently or separately downstream and the like. White/black detection is therefore critical because failure to accurately identify background pixels in the color input image adversely affects image processing downstream.
With reference now to
FIGS. 3
a
and
3
b
, one particularly acute failure mode of the single channel segmentation processing scheme will be described. As noted above, in single channel segmentation, a single channel projection vector is formed by generating a weighted vector of the three color channels into a single predetermined normalized output vector. When the plurality of color channels from the color input image are projected onto a single vector for processing, much of the color information may be lost from one or more of the color channels resulting in a single vector which essentially “looks” white to the segmentation processing unit
24
. Essentially, color information critical to proper segmentation may be lost during the pre-segmentation processing. When this happens, a pixel is labeled as belonging to a “white” or “background” class in error. Accordingly, there is a need for overriding a white classification when color content information is obliterated in the pre-segmentation processing.
Another failure mode of the single channel segmentation scheme affects white and black detection as well. With reference to
FIG. 3
a
, a color input image
30
includes a continuous color region
32
having a first color area
34
separated from a
Schweid Stuart A.
Shiau Jeng-nan
Bank One
Bayat Ali
Tran Phuoc
LandOfFree
Method and apparatus for re-classifying color image pixels... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and apparatus for re-classifying color image pixels..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and apparatus for re-classifying color image pixels... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3029800