Target-Aspect-Guided Neural Network With Geometric Constraints for Imbalanced Radar Target Recognition

Conventional high resolution range profile (HRRP) methods usually exhibit superior performance using balanced datasets but encounter severe overfitting problems when applied to imbalanced datasets. To solve this problem, a target-aspectguided neural network (TAGNet) with geometric constraints is proposed in this paper for imbalanced HRRP target recognition. In the TAGNet, based on the assumption that HRRPs can be represented by a few atoms of a scattering center dictionary (SCD), a scattering center layer (SC-layer) is designed with unrolled network architectures involving the sparse recovery optimization problem for scattering centers in HRRPs. In addition to the SCD, a discriminative dictionary with shared target-aspect atoms (DDSA) is further proposed and incorporated into the optimization problem. The SCD and DDSA are trainable and considered as parameters of the SC-layer. The DDSA not only enhances the discriminability of different target features but also improves the transferability between same target features of neighboring target-aspect frames. Leveraging the transferability, a target-aspect subprototype classifier with geometric constraints is further proposed. The proposed classifier learns target-aspect subprototypes for different targetaspect frames of each target and uses them to achieve classification through nearest neighbor search. In the classifier, the geometric constraints of majority classes are transferred to the minority classes, thus generating new target-aspect subprototypes for the minority classes. Therefore, the proposed classifier improves the performance of minority classes without the cost of performance of majority classes. In addition, a disentanglement loss is further proposed to filter out the majority-class-relevant information that possibly contained in the geometric constraints. The effectiveness and superior performance of the TAGNet for imbalanced HRRP target recognition are demonstrated through several experiments on different datasets