Scholarlink

<p>Deep convolutional neural networks (CNNs) have made tremendous progress in the object detection and classification area. In synthetic aperture radar (SAR) images, there is a need for a real-time target recognition system for onboard satellite applications. Also, the wrong classification is one of the main problems in SAR automatic target recognition (ATR) systems. So, we have introduced a lightweight SAR target classification model (less than 10&#xa0;MB size) to address these problems. Also, we developed a preprocessing technique exclusively for SAR data by utilizing the advantages of both magnitude and phase information. This preprocessing technique has the ability to reduce wrong classifications and improves the quality of&#xa0;discriminating features associated with the&#xa0;classification of SAR targets. Initially, the magnitude and phase information from the SAR patches are logically combined to generate preprocessed images. Then these images are used to train the deep learning network developed from scratch. The feature merging network and lightweight backbone network of the proposed model improve its performance further. Experimental results on the public Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset reveal that the accuracy and speed of our proposed system are both superior to the other nine state-of-the-art object classifiers.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Approach for Classifying SAR Targets Using Magnitude and Phase Information

  • 期刊论文

  • 发表日期：

    2025年8月14日

查看全文
  • C. V. Priya,
  • Anil Raj

摘要

Deep convolutional neural networks (CNNs) have made tremendous progress in the object detection and classification area. In synthetic aperture radar (SAR) images, there is a need for a real-time target recognition system for onboard satellite applications. Also, the wrong classification is one of the main problems in SAR automatic target recognition (ATR) systems. So, we have introduced a lightweight SAR target classification model (less than 10 MB size) to address these problems. Also, we developed a preprocessing technique exclusively for SAR data by utilizing the advantages of both magnitude and phase information. This preprocessing technique has the ability to reduce wrong classifications and improves the quality of discriminating features associated with the classification of SAR targets. Initially, the magnitude and phase information from the SAR patches are logically combined to generate preprocessed images. Then these images are used to train the deep learning network developed from scratch. The feature merging network and lightweight backbone network of the proposed model improve its performance further. Experimental results on the public Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset reveal that the accuracy and speed of our proposed system are both superior to the other nine state-of-the-art object classifiers.