System Analysis of VLSI Architecture for 5/3 and 1/3 Motion-Compensated Temporal Filtering

Motion-compensated temporal filtering (MCTF) is an innovative prediction scheme for video coding and it has become the core technology of the coming video coding standard, Scalable Video Coding. Since MCTF is important, this paper provides the system analysis of MCTF for hardware architecture design, including computational complexity, external memory bandwidth, and external memory size. The one-level MCTF is analyzed first, in which several frame-level data reuse schemes are proposed and the tradeoffs between external memory usages and on-chip memory size in these frame-level data reuse schemes are also discussed. Next, the analysis is extended to multilevel MCTF. The computational complexity of multilevel MCTF is close to that of traditional MC prediction with two reference frames. The memory bandwidth of multilevel MCTF depends on the frame-level data reuse scheme and performing the update stage or not. The external memory size is linearly proportional to the number of decomposition levels. Finally, a real-life test case is given to compare the system requirements between MCTF with various frame-level data reuse schemes and the prediction scheme of H.264/AVC.

I. Introduction

Existing hybrid video standards, such as MPEG series [2]–[4] and the emerging H.264/AVC [5], mainly consist of a close-loop motion-compensated prediction (MCP) scheme and a transform-based texture coder. The “close-loop” means it uses the reconstructed previous frames to predict the current frame, which forms a feedback loop. The close-loop MCP scheme has been highly optimized for the compression efficiency in the last decade, and H.264/AVC is a landmark of this development. However, for many video applications in the present and the future, the spatial, temporal, and signal-to-noise-ratio (SNR) scalabilities become more and more demanded. The scalability means we can have multiple adaptations in one video bitstream, such as different frame sizes, frame rates, and visual qualities. However, the close-loop MCP scheme is hard to provide these scalabilities while maintaining a high compression efficiency due to the drift problem, which is the mismatch of the reconstructed frame between the encoder and decoder. In order to avoid the drift problem, the compression efficiency will be degraded very much and become unacceptable when there are many scalability layers.