Image analysis – Image transformation or preprocessing – Changing the image coordinates
Reexamination Certificate
1999-07-28
2004-01-20
Mehta, Bhavesh M. (Department: 2625)
Image analysis
Image transformation or preprocessing
Changing the image coordinates
C382S260000, C348S581000
Reexamination Certificate
active
06681059
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates generally to digital video resizing or video image scaling technology and, more particularly, to techniques for reducing the cost and complexity of video scaling with minimal loss of perceived image quality.
2. Description of the Related Art
Video image scaling converts a digital or digitized image from one spatial resolution to another. For example, a digital image with a spatial resolution of 720 horizontal by 480 vertical pixels may have to be converted to another resolution in order to be displayed on a particular display device. Converting this image to, for example, a LCD panel with a fixed resolution of 640×480 requires horizontal scaling by the ratio 640/720, which is equivalent to 8/9.
This is an example of downscaling because the ratio is a fraction that is less than 1. Down-scaling creates fewer output samples than originally present in a given input. In contrast, scaling the same output to a panel of 800×600 would require horizontal scaling by 10/9 (800/720) and vertical scaling by the ratio 5/4 (600/480). These cases are examples of upscaling because the ratio is a fraction greater than 1.
Video scaling is type of digital sample rate conversion. A known technique of accomplishing video scaling is the multirate FIR (Finite Impulse Response) digital filter that achieves high quality sample rate conversion. However, this type of processing is computationally costly because it requires several multiplications and additions per output sample. When a real time processing requirement is added, the scaling function can consume a large amount of hardware resources and make it difficult to achieve high quality sample rate conversion at low cost.
FIR filters are one of two main classes of digital filters, the other being the well known IIR (Infinite Impulse Response) digital filter. Images may be thought of as signals, and like other complex signals, images typically are made up of many frequencies. High frequencies correspond to fine detail or sharpness and low frequencies correspond to smoothly or slowly changing image features.
FIG. 1
is a diagram illustrating a FIR filter
10
of the prior art. FIR filters
10
have a useful property, known as linear phase, which means that the delay through the filter is the same for all frequencies. Unequal delay results in distortion in the image, which is why FIR filters
10
are widely used in image processing applications. Linear phase results from symmetry of the filter's coefficients.
The FIR filter
10
includes shift register
12
with a series of data registers
14
, each of which is connected to a clock
16
. Each data register
14
is connected by one of a series of filter taps
20
to one of a series of multipliers
18
. The multipliers
18
are connected to an adder
22
. Data is input into the FIR filter
10
through the shift register
12
. The output of each data register
14
is coupled by one of the series of filter taps
20
to one of a set of multipliers
18
to be multiplied by a unique coefficient C
0
-C
7
. The results from each multiplier
18
are then summed by the adder
22
to produce a filtered output sample.
The number of adjacent data samples input into the FIR filter
10
is equal to the number of filter taps
20
used and is application dependent. In general, higher performance requires a larger number of adjacent samples and therefore a larger number of filter taps
20
. The multipliers
18
have coefficient symmetry because the coefficients on the left half mirror those on the right half, i.e. C
3
=C
4
; C
2
=C
5
; C
1
=C
6
; C
0
=C
7
.
As shown, FIR filter
10
has an even number of coefficients, but FIR filters may have either even or odd numbers of coefficients. FIR filter
10
may be used to implement many types of frequency responses, such as low-pass, high-pass, bandpass, etc. The type of response is determined by the number of coefficients and by the method used to calculate the coefficients. The design task for developing a low-pass FIR filter
10
is to determine the number of filter taps
20
, which is performance and application dependent, determine its cutoff frequency, and then to calculate the filter's coefficients.
There are many ways to compute the filter's coefficients. One method is known as the Windowing method. For the application of processing
8
-bit component digital video in a high quality consumer product, computing the coefficients with a Hamming window is an acceptable method. Given the number of taps and the filter's cutoff frequency, computing the coefficients for an even number of coefficients using the window method and a Hamming window can be done with the following Equation 1:
∑
i
=
1
m
c
(
i
)
=
2
fc
2
fc
π
(
i
-
1
/
2
)
*
sin
[
2
fc
π
(
i
-
1
/
2
)
]
*
{
0.54
+
0.46
cos
[
2
π
(
i
-
1
/
2
)
/
taps
]
}
m
=
taps
/
2
(
m
unique
coefficients
result
from
the
filter
'
s
symmetry
)
i
=
iteration
variable
fc
=
normalized
cutoff
frequency
(
cutoff
frquency
)
/
(
sampling
frequency
)
ranging
from
0
to
0.5
Hz
Scaling up by an integer (L) can be done directly with the FIR filter
10
. Scaling down by 1/M (M is an integer) can also be done directly with the FIR filter
10
. Video scaling typically requires scaling by a ratio of integers L/M. Scaling by a ratio is known as multirate filtering. Conceptually, it can be viewed as first upscaling by L then downscaling by M as shown in a method
24
in FIG.
2
. First, a video stream
26
is input into the FIR filter
10
. The FIR filter
10
then upscales by integer L as indicated at
28
to produce a data output by FIR filter
10
at a rate=fin*L as indicated at
30
.
Next, the FIR filter
10
downscales by integer 1/M in act
32
. This causes the video to be output at a rate=fin*L/M as shown at
34
. The FIR filter
10
accomplishes downscaling by limiting the frequency content of the input stream to less than the cutoff frequency using the low pass FIR filter
10
, then simply taking every M
th
sample and discarding the rest. After determining the number of taps, the downscaling filter's nominal normalized cutoff frequency is:
fc
=
1
2
M
.
Upscaling is more complicated. First the data stream is padded out with L−1 zero values between each input sample as shown in the example below. For L=3, if a, b, c, d, e represent a series of input data samples, the zero inserted stream becomes: a, 0, 0, b, 0, 0, c, 0, 0, d, 0, 0, e, 0, 0 . . . . This stream becomes the input to the FIR filter which is operating at a clock rate of L*fin. The padding out of zeros introduces a new frequency into the data stream, i.e. normalized introduced frequency
=
1
2
L
.
So the frequency content of the new zero padded stream consists of the original data stream plus the new frequency 1/(2L), which will always be the highest frequency in the zero padded stream. The job of the FIR filter
10
is to remove the 1/(2L) frequency and distribute the energy of the non-zero samples over all the output samples. The cutoff frequency then becomes fc=1/(2L). In addition, the energy level of the input stream must be raised by L times (because of the averaging with zero that occurs in the filter). The result is that each coefficient in Equation
1
must be multiplied by L so the coefficient calculation gives us Equation 2:
∑
i
=
1
m
c
(
i
)
=
2
Lfc
2
fc
π
(
i
DVDO, Inc.
Mehta Bhavesh M.
Patel Kanji
Perkins Coie LLP
LandOfFree
Method and apparatus for efficient video scaling 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 efficient video scaling, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and apparatus for efficient video scaling will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3262489