Rate-distortion optimized spatial scalability for DCT-based video coding

Summary form only given. We present our work on rate-distortion (RD) optimized spatial scalability for MC-DCT based video coding. Extending our work on RD optimized coding from the single layered to the multi-layered framework, we incorporate the additional inter-layer coding dependencies present in a multilayered framework into the set of permissible coding parameters. We employ the Lagrangian rate-distortion functional as it provides an elegant framework for determining the optimal choice of motion vectors, coding modes, and quantized coefficient levels by weighting a distortion term against a resulting rate term. We obtain a simple relationship between the Lagrangian parameter /spl lambda/, that controls rate-distortion tradeoffs, and the reference and enhancement layer quantization parameters QP, to allow the RD optimized framework to work easily in conjunction with rate control techniques that control the average bit rate by adjusting the quantization parameters. We then incorporate these relationships into our coder and generate two-layer bit streams with both the non-RD optimized coder and the RD optimized coder. We also generate RD optimized single-layer bit streams with the same resolution as the second layer of the two-layer bit streams. For the two-layer bit streams, we obtain a 0.6 to 1.4 dB improvement in PSNR by using RD optimization in both the base and enhancement layers. Compared to the single-layer bit stream, RD optimization in both the base and enhancement layers causes the decrease in PSNR to be reduced from 1.1 to 1.7 dB, to 0.3 to 0.5 dB.