A Target Recognition Method for High-Resolution Range Profiles in Imbalanced Datasets

Traditional high-resolution range profile (HRRP) recognition methods face challenges in automatically extracting targets’ temporal and spatial information and demonstrate low recognition accuracy on multi-class imbalanced datasets. We propose an HRRP target recognition approach based on a lightweight Transformer and fully convolutional network (FCN) model to address these challenges. The Transformer's encoder and the FCN are utilized to extract deep temporal correlation features and deep spatial structural features of HRRP, respectively. An adaptive weight matrix is introduced at the decision level to enhance the expression capabilities of temporal features and spatial features. To address the class imbalance issue, we design an adaptive focal loss function combined with a label smoothing strategy, which enhances the model's sensitivity to minority classes and improves its generalization ability. Extensive experiments conducted on the MSTAR dataset validate the effectiveness of the proposed method. The model demonstrates superior recognition performance on both balanced and imbalanced datasets, and the proposed loss function outperforms other methods.