Scholarlink

<p>The accuracy of analysis and classification in synthetic aperture radar (SAR) images is degraded by factors such as speckle noise and nonlinear energy scattering. In contrast to optical data, effective features in SAR images are derived from complex statistical structures and heterogeneous textures, which necessitates the design of purpose-built architectures for feature extraction and noise mitigation. This paper develops a hybrid deep architecture that integrates convolutional neural networks with Gabor wavelet-based attention modules. This architecture facilitates the simultaneous extraction of local features in both the spatial and frequency domains, thereby enhancing robustness against noise and variations in viewing angle. Evaluation on real SAR data indicates that the proposed method achieves a mean classification accuracy exceeding 97%. These results confirm the model’s robustness in mitigating inherent noise and spatial misalignments within SAR images, while also demonstrating the effectiveness of spatial-frequency analysis in enhancing classification performance.</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

Target recognition improvement in different angles of the SAR images based on speckle noise reduction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月15日

查看全文
  • AliAsghar Soltanali,
  • Vahid Ghods,
  • Seyed Farhood Mousavizadeh,
  • Meysam Amirahmadi

摘要

The accuracy of analysis and classification in synthetic aperture radar (SAR) images is degraded by factors such as speckle noise and nonlinear energy scattering. In contrast to optical data, effective features in SAR images are derived from complex statistical structures and heterogeneous textures, which necessitates the design of purpose-built architectures for feature extraction and noise mitigation. This paper develops a hybrid deep architecture that integrates convolutional neural networks with Gabor wavelet-based attention modules. This architecture facilitates the simultaneous extraction of local features in both the spatial and frequency domains, thereby enhancing robustness against noise and variations in viewing angle. Evaluation on real SAR data indicates that the proposed method achieves a mean classification accuracy exceeding 97%. These results confirm the model’s robustness in mitigating inherent noise and spatial misalignments within SAR images, while also demonstrating the effectiveness of spatial-frequency analysis in enhancing classification performance.