A Deep Learning-Based SAR Imaging Framework for Ship Targets With Sample-Wise Variant Motion

Obtaining the clear contours of ship targets via synthetic aperture radar (SAR) is extremely valuable for monitoring the sea. Now there are many deep learning imaging methods for ground scenes with good results, but they will face three main challenges when imaging ship targets: 1) the translational and rotational motions of ship targets during travel and due to waves, respectively, bring spatial invariant and variant errors that are tough to estimate and compensate, resulting in defocused SAR imaging results; 2) the varying motion of ships demands a high generalization ability of the imaging reconstruction network to adapt to the ship targets with sample-wise variant motion parameters; and 3) since ships are noncooperative targets, the accuracy of motion parameter estimation should be evaluated based on image quality, whereas the available SAR image quality assessment (IQA) functions exhibit limited robustness. To address these issues, this article proposes a deep learning-based SAR imaging framework for ship targets via deep unfolding. First, the motion model and characteristics of ship targets are analyzed, and the SAR imaging model for ship targets with complex translational and rotational motion is established. Second, an imaging network with high generalization ability is proposed to adapt echoes for imaging under different motion parameters. On this basis, an SAR ship IQA network is proposed to assess the imaging results of SAR ship targets with different focusing qualities. Then, the high-resolution imaging problem of ship targets is regarded as a motion parameter optimization problem, with the IQA results as the objective function. Finally, this problem is optimized to search for the most accurate translational and rotational motion parameter variables of the ship target, achieving error compensation and imaging. The validity of the method is verified through the simulation of point targets and real SAR scenario data.