A Dual-Driven Deep Learning Model With Local-Structure Physical Scattering Characteristics and Fine-Grained Image Features for SAR Ship Recognition

In recent years, synthetic aperture radar (SAR) ship recognition based on deep learning has received extensive attention and research. The accuracy and reliability of the recognition model can be effectively improved by fusing physical scattering characteristics and image features, which has become one of the important trends in SAR recognition research. However, there are some challenges in feature extraction and fusion of the scattering characteristics and image features. To address these issues, a new dual-driven deep learning model with local-structure physical scattering characteristics and fine-grained image features (DDLM) is proposed. First, an integrated Local and Global physical Scattering characteristics Aware Module (LGSAM) is constructed. The spatial information and structural attributes of the sets of attribute scattering centers (ASCs) are co-feature encoded to obtain local and global structural features of ASCs in the scattering domain. Second, an Adaptive Multi-scale feature Refinement extraction Module (AMRM) is designed. The AMRM can extract multi-scale fine-grained features and focus more on the target, thereby effectively reducing interference from background noise in the image domain. Third, based on the significant differences in the information and structure contained in the scattering characteristics and image features, a Differential Feature Fusion of semantic alignment Module (DFFM) is developed. The DFFM can effectively fuse the differential features extracted, fully utilize the complementarity between the two features, and strengthen the learning ability as well as the interpretability of the model. Finally, experimental validation is carried out using the ComplexSAR_Ship and the MSTAR dataset. The DDLM achieves Accuracy, Precision, Recall, and F1 of 95.32, 89.30, 89.45, and 89.28, respectively, in ComplexSAR_Ship.