Bayesian Image and Video Enhancement Using a Non-Gaussian Prior - Nonlinear Signal and Image Processing: Theory, Methods and Applications - page 325

(e) (f)

FIGURE 9.15
( Continued )

an IEEE 1394 (also known as FireWire) connection and the DV-NTSC standard
within Adobe Premiere (720 480 pixel frames; 29.97 fps frame rate; standard

NTSC video with a frame aspect ratio of 4:3 and interlacing; pixel aspect ratio
of 0.9). Interlacing, which further reduces the number of useful constraints
available for enhancing a particular video frame, is incorporated into the
video observation model through H ( i ) . In terms of appearance, the linear
(Gauss-Markov) estimate in Figure 9.18 is more aesthetically pleasing than the
nonlinear (Huber-Markov) estimate, probably because the jagged structures
that are present due to the lack of constraints available in alternating scan
lines become smoothed in the linear estimate. Note that the motion confidence
parameter was further reduced to

×

λ ( i ) = 2 in this real video example, because

of the nonuniformity of motion across the scenes. In the simulation results
depicted in Figure 9.18, no downsampling of the original data took place prior to
super-resolution enhancement. Thus, unlike most simulation results presented in
the literature, this experiment is representative of real-world superresolution
enhancement.
As the world moves farther in the direction of digital communications, low
bit rate video compression is becoming increasingly important, 29 , 30 and su-
perresolution enhancement can help restore some of the information lost due
to encoding. However, the success of superresolution enhancement is pred-
icated on the redundancy of the data found within a video sequence. When
that redundancy is eliminated through intraframe and interframe compres-
sion, it becomes almost futile to extract details from an arbitrary sequence. 16
Superresolution enhancement has been applied to MPEG sequences (high
variable compression ratios in excess of 10:1) in the past, 14 - 16 although so
much redundant data is removed in the interframe coding process that the