Resolution enhancement for inverse synthetic aperture radar images using a deep residual network

A framework for inverse synthetic aperture radar (ISAR) image resolution enhancement using a deep residual network is proposed. Our framework directly learns an end-to-end mapping in the form of a deep residual network (ResNet) between the input low-resolution (LR) images and the output high-resolution (HR) images with respect to the point spread function (PSF). In our network, residual blocks without batch normalization (BN) layers are applied to retain ISAR image contrast, and parametric rectified linear unit (PReLU) which can adaptively learn the negative coefficients is used as the activation function. The complex-valued ISAR imagery is divided into two real-valued channels to preserve phase information implicitly. Experimental results show that the proposed method can obtain higher quality HR images compared with traditional sparsity-driven methods while freeing itself from a prior model for radar echoes and a complicated parameter estimation process.