ADS Technologies USBAV-708 INSTANT VIDEOMPX, User Manual

Page: 57 / 116

The
Guide
57
Ps
yc
ho
vis
ua
l E
nha
nc
em
en
t
Th
e D
C
T,
iD
C
T
a
nd
H
um
a
n
Vis
ua
l S
ys
te
m
,
Th
e p
sy
ch
ovi
sua
l e
nh
a
nc
em
ent
sy
ste
m
The human visual system is far less sensitive to high frequencies in an image
than it is to low frequencies. One perceptual technique DivX uses during lossy
compression is the reduction in accuracy of higher frequency co-efficients, saving
bits while causing the least perceivable quality degradation.
It is in fact these DCT co-efficients that are quantized by DivX when image data
is encoded (see
Forward—Quantizers) . Greater quantization of the DCT co-
efficients means fewer bits are ultimately required to store them, but a less
accurate image results when the
inverse discrete cosine transformation (or iDCT)
is performed during decoding—the process of restoring the original series (image
data) from the DCT result. This error between source and encoded image is
known as
quantization noise .
As described, image data is stored by DivX as a set of co-efficients linked to
different frequencies in the image. The psychovisual enhancement process
actually manipulates these co-efficients to produce two distinct effects:
In flat areas of the image co-efficients are manipulated so that fine details
are enhanced. If we were to encode in 1-Pass Quality Based mode (i.e. at
a fixed quantizer) this enhancement would naturally
increase the bits spent
on flat areas of the image. However, at a fixed average bitrate the effect is
actually that textured areas of the image will receive fewer bits and hence
when psycho-visual enhancements are enabled artifacts will be masked in
textured areas of the image where they are least visible.
This works because of a balancing effect
that is created - the encoder has a finite
number of bits to spend and if flat areas are
to consume more bits then textured areas
must consume fewer bits.
In strongly textured areas of the image co-efficients are manipulated so
that fewer bits are spent and thus quality is slightly degraded. If we were to
encode in 1-Pass Quality Based mode (i.e. at a fixed quantizer) this
enhancement would naturally decrease the bits spent on heavily textured
areas of the image. However, at a
fixed average bitrate the effect is
actually that flatter areas of the image will receive more bits (once again
due to balancing) and hence when psycho-visual enhancements are
enabled artifacts will be masked in textured areas of the image where they
are least visible.
The psychovisual enhancement system
Flat Textured
Original bits
Bits after PV
1 .
2 .