摘要
相位恢复问题是指仅从幅值测量中恢复原始信号.由于幅值测量中缺少相位信息,精确恢复原始信号困难,因此需要加入正则化项确保高精度重建原始信号.结合交替投影和卷积神经网络提出了基于卷积神经网络去噪正则化的相位恢复算法(NrPR_DnCNN).所提算法将相位恢复问题转化为去噪和约束优化两个子问题,并利用l 1正则化快速梯度下降法交替求解.仿真结果表明:与BM3D_PRGAMP算法相比,所提算法重构图像的峰值信噪比在二种高斯噪声水平上分别提高了2.08 dB和3.20 dB,验证了所提算法的有效性和鲁棒性;误差-迭代仿真结果验证了所提算法具有良好的收敛性.
The phase retrieval problem is to recover the original signal only from the magnitude measurement.Due to the lack of phase information in the magnitude measurement,it is difficult to recover the original signal accurately.Therefore,it is necessary to add regularization term to ensure the reconstruction of the original signal with high accuracy.In this paper,a phase retrieval algorithm(NrPR_DnCNN)based on denoising regularization of convolutional neural network is proposed by combining alternating projection and convolutional neural network.The proposed algorithm transforms the phase retrieval problem into two subproblems:denoising and constrained optimization.This problem is solved alternately by using a fast gradient descent based on the l 1 regularization.The simulation results show that:compared with BM3D_PRGAMP algorithm,the peak signal to noise ratio of the image reconstructed by the proposed algorithm is increased by 2.08 dB and 3.20 dB at two Gaussian noise levels,respectively.This verifies the effectiveness and robustness of the proposed algorithm.The error-iteration simulation results indicate that the proposed algorithm has good convergence.
作者
项宇
李岚
蒲莎莎
XIANG Yu;LI Lan;PU Sha-sha(School of Science,Xi'an Shiyou University,Xian 710065,China)
出处
《西安文理学院学报（自然科学版）》
2023年第2期22-28,共7页
Journal of Xi’an University(Natural Science Edition)
基金
陕西省自然科学基础研究计划资助项目(2021JM-399)
西安石油大学研究生创新基金项目(YCS21212152)。
关键词
相位恢复
卷积神经网络
去噪
交替投影
正则化项
快速梯度下降法
phase retrieval
convolutional neural network
denoising
alternating projection
regularization term
fast gradient descent method
