An Efficient Network Based on Conjugate Gradient Optimization and Approximate Observation Model for SAR Image Reconstruction

Deep learning has been successfully applied to solve the synthetic aperture radar (SAR) imaging problem, which shows superior imaging performance to compressive sensing (CS)-based methods under sparse sampling conditions. However, due to the computation of the large-scale matrix, the optimal searching in an iterative manner will involve tremendous computational complexity with slow convergence, which will prevent deep learning algorithms from being efficiently applied for SAR imaging. To address this problem, an efficient network is proposed for SAR imaging under sparse sampling conditions, which can be designed by an improved conjugate gradient (CG) optimization strategy. First, the large-scale matrix in the CG algorithm can be approximately decomposed by introducing a matched filtering (MF)-based operator, which will facilitate the gradient computation with high efficiency during the optimization process. Second, the strategy utilizes the advantages of CG optimization to precisely eliminate the error component with conjugate searching in each iteration to achieve fast convergence. By incorporating the improved CG into the convolutional neural network (CNN), the network can be developed to automatically learn the prior information and parameters from the training data, based on which the efficiency of the designed network can be increased dramatically to achieve high imaging performance for SAR applications, especially in the cases of wide-scene and high-resolution imaging. Experimental results show that the proposed network exhibits excellent imaging performance and high computational efficiency in SAR imaging of point targets, surface targets, and different types of real scenes.