Encoder optimization in an extended H.263 framework

In a motion-compensated discrete-cosine-transform (MC/DCT) coding framework similar to the H.263-standard video coder, we explore the performance gain furnished by rate-distortion (R-D) optimization. We show that exact global R-D optimization is a very complex task. We seek good performance-complexity tradeoffs instead. We explore an encoding scheme that has several complexity reduction features: noniterative Lagrangian motion estimation, optimized table-based estimators of DCT coding bit rate and distortion performance, and bottom-up propagation of block matching results. Our scheme is compared with the advanced prediction mode (APM) of H.263, as instrumented by the popular Telenor test model. For typical QCIF video test sequences, our scheme furnishes PSNR gains ranging from 0.35 to 1.1 dB. Visual quality improvement is highly palpable. The complexity of our scheme is roughly 30% higher than the Telenor implementation.