MMSE准则下基于玻尔兹曼机的快速重构算法
发布时间：2018-08-22 11:53
【摘要】：全连接的玻尔兹曼机模型可全面描述稀疏系数间统计依赖关系,但时间复杂度较高.为了提高基于玻尔兹曼机的贝叶斯匹配追踪算法(BM-BMP)的重构速度和质量,本文提出一种改进算法.第一,将BM-BMP算法的最大后验概率(MAP)估计评估值分解为上一次迭代的评估值与增量,使得每次迭代仅需计算增量,极大缩短了计算耗时.第二,利用显著最大后验概率估计值平均的方式,有效近似最小均方误差(MMSE)估计,获得了更小的重构误差.实验结果表明,本文算法比BM-BMP算法的运行时间平均缩短了73.66%,峰值信噪比(PSNR)值平均提高了0.57 d B.
[Abstract]:The fully connected Boltzmann machine model can fully describe the statistical dependence between sparse coefficients, but the time complexity is high. In order to improve the reconstruction speed and quality of Bayesian matching tracking algorithm (BM-BMP) based on Boltzmann machine, an improved algorithm is proposed in this paper. Firstly, the maximum posterior probability (MAP) estimation of BM-BMP algorithm is decomposed into the evaluation value and increment of the previous iteration, which makes each iteration only need to calculate the increment, which greatly shortens the computation time. Secondly, the minimum mean square error (MMSE) estimation is effectively approximated by the average value of the significant maximum posterior probability, and a smaller reconstruction error is obtained. The experimental results show that the average running time of this algorithm is 73.66 shorter than that of BM-BMP algorithm, and the (PSNR) value of peak signal-to-noise ratio is increased by 0.57 dB on average.
【作者单位】： 华南理工大学电子与信息学院;国家移动超声探测工程技术研究中心;
【基金】：国家自然科学基金资助项目(61327005,61302120) 广东省科技计划资助项目(2017A020214011) 中央高校基本科研业务费资助项目(2017MS039)
【分类号】：TN911.7
本文编号：2197018
[Abstract]:The fully connected Boltzmann machine model can fully describe the statistical dependence between sparse coefficients, but the time complexity is high. In order to improve the reconstruction speed and quality of Bayesian matching tracking algorithm (BM-BMP) based on Boltzmann machine, an improved algorithm is proposed in this paper. Firstly, the maximum posterior probability (MAP) estimation of BM-BMP algorithm is decomposed into the evaluation value and increment of the previous iteration, which makes each iteration only need to calculate the increment, which greatly shortens the computation time. Secondly, the minimum mean square error (MMSE) estimation is effectively approximated by the average value of the significant maximum posterior probability, and a smaller reconstruction error is obtained. The experimental results show that the average running time of this algorithm is 73.66 shorter than that of BM-BMP algorithm, and the (PSNR) value of peak signal-to-noise ratio is increased by 0.57 dB on average.
【作者单位】： 华南理工大学电子与信息学院;国家移动超声探测工程技术研究中心;
【基金】：国家自然科学基金资助项目(61327005,61302120) 广东省科技计划资助项目(2017A020214011) 中央高校基本科研业务费资助项目(2017MS039)
【分类号】：TN911.7
【相似文献】
相关硕士学位论文 前1条
1 郭岩河;基于深度学习的极化SAR分类研究[D];西安电子科技大学;2015年
，本文编号：2197018
本文链接：https://www.wllwen.com/kejilunwen/xinxigongchenglunwen/2197018.html
