SAR Super-Resolution Imaging and Recognition Integrated Network Based on Deep Learning Framework

This is the authors' abstract. We don't add key points for this paper.

Deep learning (DL) has been widely applied in synthetic aperture radar (SAR) imaging and recognition technology, including multiple operations such as image formation, enhancement, and recognition. Currently, DL methods are only utilized in conjunction with individual image processing operations to enhance corresponding computational performance. The limitations of resolution and data storage also pose significant challenges to imaging and recognition. These limit SAR's ability to fully exploit the advantages of DL and deliver accurate and fast responses from echo signals to target categories. This article proposes a DL framework-based super-resolution imaging (SRIm) and recognition integrated method, called DL framework-based integrated network (DLIN). This method can simultaneously complete echo imaging and target recognition tasks, that is, to achieve SRIm from low data echoes and complete fast recognition. Specifically, DLIN is a novel SAR processing pipeline that includes a SRIm network (SRIm-Net), followed by a Transformer-based recognition network (Recog-Net). SRIm-Net first uses an iterative shrinkage threshold algorithm (ISTA)-based 2D deep unfolding network (2D-DUN) to achieve sparse imaging of echo signals. Then, it introduces a neural network architecture called image super-resolution network (SR-Net), which integrates ResNet and U-Net to obtain SRIm results. Finally, Recog-Net is used to complete the image classification task. DLIN can obtain real-time processing results of echo data, including highquality super-resolution images and accurate target categories. The effectiveness of our proposed DLIN is validated through simulated and measured experiments. The experimental results have shown that DLIN can achieve SRIm, and its improved imaging resolution further enhances the recognition accuracy.