SAR Target Recognition Network Based on Time-Frequency Domain Channel Attention Mechanism

This paper introduces a Synthetic Aperture Radar (SAR) target recognition network based on the time-frequency domain channel attention mechanism (DSN-DCT). Traditional SAR target recognition networks utilize the ResNet model to extract time-domain and frequency-domain features from SAR images. However, these networks suffer from low utilization rates of frequency-domain channels, leading to insufficient extraction of feature information and relatively low recognition accuracy. To address this issue, we have integrated a Discrete Cosine Transform (DCT) based time-frequency domain channel attention mechanism module into the existing deep network architecture. This module is capable of extracting the most significant top N high-frequency components in the spectral space, effectively capturing detailed information such as edges and textures in SAR images, which is crucial for a deeper understanding of SAR imagery. Furthermore, considering the overfitting issues associated with DCT when applied to small datasets, and to ensure that the model maintains high recognition accuracy in the presence of complex noise, we employ data augmentation techniques. Specifically, we add Gaussian noise to SAR images and integrate the noisy image features with the corresponding generated spectral features. These enhanced features are then input into the target recognition network for training and prediction. Ultimately, on the Sentinel-1 dataset, we have improved the accuracy from 91.25% to 93.44%, and on the OpenSARShip_1 dataset, we have increased the accuracy from 72.80% to 76.60%. These results substantiate the effectiveness and generalizability of our proposed method.