Sequential ISAR Images Classification Using CNN-Bi-LSTM Method

The time-varying radar line of sight (LOS), and the changing range and azimuth sampling rates bring unknow deformation to inverse synthetic aperture radar (ISAR) imaging, such as rotation and scaling, which makes space target classification more difficult. To effectively classify ISAR images with deformation, a sequential ISAR images classification method combining CNN and bidirectional long short-term memory (CNN-Bi-LSTM) is proposed in this paper. In this structure, Bi-LSTM can fuse the features of single ISAR image extracted by CNN, and then extract deformation-robust features of ISAR images. For the ISAR image dataset generated from electromagnetic (EM) echoes of satellite targets, the proposed method achieves higher classification accuracy than traditional methods.