کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

SRNet-2D: A Sparse-Representation Deep Network for Robust SAR Target Recognition

نویسندگان :

Fatemeh Eslami¹ Sadegh Samadi² Reza Mohseni³

1- دانشگاه صنعتی شیراز 2- دانشگاه صنعتی شیراز 3- دانشگاه صنعتی شیراز

کلمات کلیدی :

Sparse Representation،SRNet-2D،Automatic Target Recognition (ATR)،Synthetic Aperture Radar (SAR)،MSTAR Dataset،Real and Synthetic SAR Imagery

چکیده :

Sparse-representation deep network modeling has been effectively used for enhancing features in one-dimensional vibration signals, however its application to two-dimensional signals has been limited. This paper introduces SRNet-2D, a deep architecture that combines a two-dimensional sparse-representation layer with selective residual learning for automatic target recognition (ATR) in synthetic aperture radar (SAR) imagery. The SR layer learns a dictionary of convolutional atoms and estimates sparse activation coefficients to highlight important scattering structures while minimizing background noise. Experiments on the MSTAR dataset demonstrate the effectiveness of SRNet-2D. The model achieves 92.62% accuracy on the REAL subset and 92% accuracy on the SYNTH subset, where most misclassifications are due to differences in synthetic and real-world scattering patterns. SRNet-2D provides a lightweight, interpretable, and effective solution for SAR target recognition, combining sparse modeling with deep learning.