An Efficient Mode Decision Algorithm for Combined Scalable Video Coding - Advanced Concepts for Intelligent Vision Systems - page 66

Table 3. Simulation results for combined scalability

Δ Time(sec)

BDPSNR (dB) BDBR (%)

Sequence

Li's Ours Li's Ours Li's Ours

FOREMAN -0.67 -0.07 14.79 1.60 39.25 48.38

MOBILE -0.31 -0.05 6.90 1.33 38.94 43.82

CITY -0.71 -0.04 13.51 0.89 38.32 47.68

BUS -0.52 -0.10 9.53 1.80 39.13 46.49

SOCCER -0.61 -0.08 11.07 1.50 38.85 47.94

FOOTBALL -0.56 -0.23 9.48 3.90 36.91 44.11

ICE -0.70 -0.05 17.20 1.50 37.89 46.98

HARBOUR -0.27 -0.03 7.60 1.00 40.52 43.10

Average -0.54 -0.08 11.26 1.69 38.72 46.06

Time = Time [ reference ]

Time [ P roposed ]
Time [ reference ]

−

× 100 (8)

We used Li's method [10], a well known fast mode decision technique in the SVC
encoding system, for an objective comparison of the encoding performance of our
algorithm to provide spatial and temporal scalability. Results for Li's method
are shown separately [10]. For a comparison with our algorithm, we implemented
Li's method to support combined scalability because our proposed algorithm is
designed to support combined scalability.
Table 3 shows the simulation results for combined scalability with various
QP values. The average loss in BDPSNR was measured as -0.23 ∼ 0.03dB and

BDBR increased 0.89 3.9%, compared with the full mode search. The proposed

algorithm increases the speed of the SVC encoding system up to 48.38% at
FOREMAN, compared to the full mode search. Compared to Li's method, the
proposed algorithm achieved a speed-up gain of up to 9% with a smaller bit
increment. Li's method resulted in a large quality loss (0.54 (dB)) and a large
bit increment (11.26%) for combined scalability. The proposed algorithm resulted
in a speed-up gain of approximately 8% more than Li's method while suffering
less quality loss and a smaller bit rate increment.

∼

5 Conclusions

A fast mode decision algorithm is proposed for inter-frame encoding with com-
bined scalability. This algorithm is based on correlative information between the
base layer and enhancement layers and correlation of temporal levels. In the
proposed algorithm, we define a cost for the motion area using ordered mode in-
formation. Our scheme also uses two classes for the mode search and a feedback
structure to guarantee image quality. For combined scalability, experimental re-
sults show that the proposed algorithm significantly reduces the computational
complexity of the SVC encoder up to 48% with only a small PSNR loss and bit
rate increment.