MMFF: Multiview and multi-level feature fusion method within limited sample conditions for SAR image target recognition

Abstract

The fusion of SAR image features from multiple views can effectively improve the recognition performance of SAR ATR tasks. However, when the number of raw samples in SAR images is limited, multiple fusions of SAR image features from different views of the same class may result in significant feature redundancy, causing overfitting of the model. To solve those problems, we propose a multiview and multi-level feature fusion (MMFF) method that can extract richer features from extremely limited raw data. Firstly, we design a new multiview feature fusion (NMFF) module to reduce feature redundancy generated by fusing features from the same class but from different views. This module uses multiple feature fusion methods to fuse features from different views, effectively reducing feature redundancy and alleviating model overfitting. Then, we design a multiview multi-class random feature extraction (MMRFE) module to extract inter-class separability features and intra-class similarity features and fuse them with multiview features. The MMRFE module enables the network to learn inter-class separability between different classes and intra-class similarity between the same classes, thereby improving the network’s recognition ability in extremely limited data. Finally, to further increase inter-class separability and intra-class similarity, we design a coarse classifier to perform coarse classification on inter-class separability features and intra-class similarity features. The coarse classifier increases inter-class separability and intra-class similarity by calculating classification loss to affect updating network parameters. Experimental results demonstrate that when trained with 10 SAR images per class, our algorithm achieves recognition rates of 92.53 % and 80.50 % on the MSTAR dataset and Civilian Vehicle dataset, respectively, outperforming state-of-the-art methods by at least 3.2 % and 3.94 % in classification accuracy.