.png){kind=icon}
.png){kind=icon}
Return
Memory-Centered Motion Estimation System With CTB-Based Full-Splitting Algorithm
.png){kind=icon}
PRE
AI
DOI：10.1109/TCE.2024.3399123.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The processing of Motion Estimation (ME) in current video encoders demands high memory communication due to the variety of supported block sizes and the high number of candidate blocks that must be evaluated. The intense memory access demanded to process UHD 4K videos in real-time results in undesirable high-power dissipation, which is notably problematic for battery-powered consumer devices. Current ME hardware designs in literature employ different strategies to reduce memory communication. However, those solutions process different blocks independently, disregarding all the potential for data and operations reuse. This paper presents a memory-centered ME hardware system targeting UHD 4K videos at 120fps featuring (i) a novel CTB-Based Full-Splitting algorithm able to exploit full data reuse across distinct block sizes, (ii) an on-chip memory organization tailored for reducing power dissipation, and (iii) a dedicated processing unit designed to reuse data and operations. Unfortunately, the power results of the global ME system, including the memory subsystem, are not properly presented in related works, so a full comparison is not possible. Nevertheless, when considering only the processing unit, our solution reached the smallest power dissipation when compared with related works.
Keywords:
Hardware
Memory management
Encoding
Streaming media
Bandwidth
Organizations
Complexity theory
Video coding
motion estimation
hardware design
data and operations reuse
memory-centered design
Journal
IF:
10.9
Papers:
5.1K
Citations:
6.8K
