A Convolutional Neural Network Using Multi-Level Feature Projection for SAR Vehicle Recognition

In synthetic aperture radar (SAR) target recognition scenarios, traditional convolutional neural network (CNN) models often struggle with inadequate generalization performance. To address the challenging operating conditions in SAR vehicle target recognition, this paper proposes a new CNN model that employs multi-level feature projection. This model first uses a lightweight backbone network to extract multi-level feature maps. These feature maps are normalized to balance their weights across different levels and channels. Then, a new feature projection module is proposed to perceive and fuse multi-level features, which addresses high-dimensional characteristics and overfitting effects. Extensive experiments on the benchmark dataset validate the effectiveness of the proposed model.