Accurate and Efficient Matrix Completion Using Cascaded Deep Neural Network

The matrix completion problem, which recover the missing data from the observed ones, had been widely studied in recent years. Although deep learning techniques had been applied in varies fields, limited works had done on matrix recovery. In this paper, we proposed a new deep neural network (DNN) model by integrating optimization theory and deep learning technique to solve the matrix completion problem. A cascaded neural network that contains the idea of alternating optimization is trained, and the application of SAR data reconstruction and imaging is used for evaluation. Experimental results shown that the proposed model can achieve better performance with less computational complexity when the sampling rate is sufficiently low.