Transformer-in-Transformer Network with A Complex-Valued Mixup Module for SAR Automatic Target Recognition

Convolutional neural network (CNN) has been demonstrated an effective approach for synthetic aperture radar (SAR) automatic target recognition (ATR) in recent years. However, CNNs usually extract characteristics through local receptive fields, resulting in the loss of global feature information of targets. Moreover, satisfactory target recognition performance using CNN generally depends on sufficient labeled sample images, which are difficult to obtain in SAR ATR conditions. To address these issues, we propose a Complex-Valued Mixup (CVM) module- aided Transformer-in-Transformer (TNT) network (CVM-TNT) for SAR ATR. The CVM module generates new samples through linear interpolation and utilizes both amplitude and phase information to extract richer electromagnetic scattering feature. The TNT network takes the advantages of global and local features to compensate the deficiency of traditional Transformer network that only extracts global features. Firstly, we divide the input images including amplitude and phase into several local patches as "visual sentences". Then, we use the outer conventional Transformer block to calculate global features and their global attentions of visual sentences. Subsequently, dividing the local patches into sub-patches as "visual words" and embedding an inner sub-Transformer block into the network excavates the local phase and amplitude features of visual words. Finally, both word-level and sentence-level features will be combined to boost the representation capability. Experiments implemented on the moving and stationary target acquisition and recognition (MSTAR) dataset have shown that the proposed CVM-TNT model outperforms the state-of-the-art (SOTA) methods in the case of small datasets.