HRPnet: High-Dimensional Feature Mapping for Radar Space Target Recognition

Deep learning has made significant progress in the field of radar space target recognition. However, deep neural networks require significant amounts of data to train network parameters, posing challenges in achieving noncooperative target recognition. Therefore, it is of practical significance to research a fast and accurate target recognition method with limited radar data. In this article, we propose a novel radar space target recognition network based on high-dimensional feature maps, called HRPnet, which fully utilizes high-resolution range profile (HRRP), radar cross section (RCS), and polarization (POL) data obtained from the radar. First, a sparse autoencoder (SAE) is employed to conduct deep feature extraction on the three types of data. Second, the Gramian angular field (GAF) transformation is employed to obtain 2-D representation of HRRP, RCS, and POL data, respectively. These 2-D maps are then integrated to construct high-dimensional feature maps. Third, a feature map convolutional neural network (FMCNN) is designed for high-dimensional feature map classification and target recognition. Experimental results indicate that the proposed HRPnet outperforms existing methods in terms of recognition accuracy and noise resistance, particularly in the case of a limited sample size.