ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2019, Vol. 56 ›› Issue (11): 2469-2484.doi: 10.7544/issn1000-1239.2019.20180699

• 图形图像 • 上一篇    下一篇

采用2 b深度像素的弹性运动估计算法

宋传鸣1,2,4,闵新1,3,谢维冬1,3,尹宝才2,王相海1   

  1. 1(辽宁师范大学计算机与信息技术学院 辽宁大连 116029);2(大连理工大学计算机科学与技术学院 辽宁大连 116024);3(东北大学计算机科学与工程学院 沈阳 110169);4(计算机软件新技术国家重点实验室(南京大学) 南京 210023) (chmsong@lnnu.edu.cn)
  • 出版日期: 2019-11-12
  • 基金资助: 
    国家自然科学基金项目(61402214,41671439,61632006);辽宁省自然科学基金项目(20180550570);辽宁省高等学校创新人才支持计划项目(64);南京大学计算机软件新技术国家重点实验室开放课题项目(KFKT2018B07)

Elastic Motion Estimation Algorithm Using Two-Bit-Depth Pixels

Song Chuanming1,2,4, Min Xin1,3, Xie Weidong1,3, Yin Baocai2, Wang Xianghai1   

  1. 1(School of Computer and Information Technology, Liaoning Normal University, Dalian, Liaoning 116029);2(School of Computer Science and Technology, Dalian University of Technology, Dalian, Liaoning 116024);3(School of Computer Science and Engineering, Northeastern University, Shenyang 110169);4(State Key Laboratory for Novel Software Technology (Nanjing University), Nanjing 210023)
  • Online: 2019-11-12

摘要: 为降低传统弹性运动估计的计算复杂度,提出一种2 b深度像素的弹性运动估计方法.首先,利用Prewitt算子提取视频帧的梯度,借助梯度模长的均值和标准差将像素深度从8 b降采样为2 b.其次,引进基于位操作的矩阵乘法和基于比较操作的偏导运算,提出了2 b深度像素的弹性运动模型以及求解该模型的简化高斯-牛顿法,避免了黑塞矩阵及其逆矩阵的重复计算.同时,采用1阶线性逼近,得出阻尼步长与运动向量增量、运动补偿误差之间的函数关系以及初始步长的近似计算策略,进而以菱形搜索为初始搜索,给出了2 b深度像素的弹性运动模型的快速求解算法.实验表明:该算法的平均峰值信噪比和计算效率明显优于8 b全搜索、2 b全搜索和传统8 b弹性运动估计.

关键词: 视频编码, 运动估计, 低位深度像素, 弹性运动模型, 自适应阻尼系数

Abstract: To reduce the computational complexity of traditional elastic motion estimation, this paper proposes a novel elastic motion estimation algorithm using two-bit-depth pixels. First, the Prewitt operator is employed to calculate the gradient of each video frame. The mean and standard deviation of the gradient norm is utilized to down-sample each pixels depth from 8 b into 2 b. Second, the bitwise operation-based matrix multiplication and the comparison based partial derivative computation are introduced. We subsequently describe an elastic motion model using two-bit-depth pixels, as well as a simplified Gaussian-Newton method which avoids the repetitive computation of the Hessian matrix and its inverse matrix. Meanwhile, we establish the functional relationship of the damping step size versus motion vector increment and motion-compensated errors by the first-order linear approximation, obtaining a method for approximately calculating the initial step size. Furthermore, we address a fast method for solving the elastic motion model with two-bit-depth pixels, using the diamond search algorithm as initial search. Experimental results illustrate that our algorithm obviously outperforms the full search with eight-bit-depth pixels, the full search with two-bit-depth pixels, as well as the conventional elastic motion estimation with eight-bit-depth pixels in terms of the peak signal-to-noise ratio (PSNR) and computational efficiency.

Key words: video coding, motion estimation, low bit-depth pixel, elastic motion model, adaptive damping coefficient

中图分类号: