Digital Signal Processing Reference
In-Depth Information
Fig. 8
Illustration of Discrete Cosine Transform (
a
) The original Lena image; (
b
) After transform,
energy is compacted to low-frequency components (the
upper-left side
).
White and black pixels
are
with high energy.
Gray pixels
are with low energy
×
N
2-D DCT
cos
2
Π
(
2
x
+
1
)
u
4
N
cos
2
Π
(
2
y
+
1
)
v
4
N
2
N
C
N
−
1
N
−
1
F
(
u
,
v
)=
(
u
)
C
(
v
)
∑
0
f
(
x
,
y
)
∑
x
=
y
=
0
1
√
2
,
(4)
if u,v = 0
C
(
u
)
,
C
(
v
)=
1
,
otherwise
F
is the pixel value at the
location (x,y). The first and the second cosine terms are the vertical/horizontal
cosine basis function. With DCT, energy will be gathered to the low-frequency parts
efficiency in entropy coding.
information in a block and leads to lossy coding. A simple example of uniform
All input value between two decision boundaries are output as the same value. For
example, the output value is
y
1
if the input value is inside the interval
x
1
<
(
u
,
v
)
is the DCT value at the location (u,v) and
f
(
x
,
y
)
x
2
.
In this example,
y
1
is set as the average value of
x
1
and
x
2
. Quantization step is the
distance between two decision boundaries. In a video coding system, quantization
step is adjusted to control the data rate. Larger quantization step leads to lower data
rate but poorer visual quality, and vice versa. In addition, human visual system is less
sensitive to high-frequency components. High-frequency components are usually
quantized more heavily with less visual quality degradation. This is another example
to remove perceptual redundancy for data compression.
X
<