面向HEVC的运动估计快速算法和硬件架构

Fast algorithm and hardware architecture for HEVC motion estimation

提出确定共享搜索区域的方法,实现硬件搜索区域共享,能有效提高数据复用和减少参考像素带宽.进而,提出改进的钻石搜索算法,该算法考虑了硬件资源消耗和预测单元大小,使预测单元自适应选择搜索模板.最后,基于改进的钻石搜索算法提出新的硬件架构,该架构通过灵活选择不同处理单元(PE)数目,实现两种基本处理单元,使不同尺寸预测单元都有较高的处理速度和硬件资源利用率.算法仿真结果表明,本算法与参考代码HM16.7相比较,编码性能损失可忽略不计,但更适合于硬件实现.用Altera的Stratix Ⅳ系列芯片在QUARTUS Ⅱ中综合结果表明,周期数比现有文献更少,本架构最大工作频率可达到317.56 MHz,并且实现1080 px@23.7帧·s^(-1)的吞吐率.

This paper proposes a method to share a fixed hardware search region which can reuse data effectively and reduce bandwidth of the reference pixel at the same time. Secondly,an improved diamond search algorithm is proposed. The algorithm considers the hardware resource consumption and the size of the prediction unit. And the prediction unit adaptively selects the search template. Finally,this paper further proposes a new hardware architecture based on improved diamond search algorithm. This hardware architecture achieves two basic process units by flexibly selecting the number of process element （ PE ）. So that different sizes of prediction units have high processing speed and hardware resource utilization. The proposed algorithm is more suitable for hardware implementation with negligible loss compared with the reference code HM16.7. The architecture has been synthesized for an Altera Stratix IV FPGA in QUARTUS II. The architecture has less the number of cycles in comparison to related works. For the architecture reaches a maximum frequency of 317.56MHz,it is possible to achieve 1 080px @23.7f·s-1.