EffiSARNet: A Lightweight Gated Multi-Stream Network for Efficient SAR Automatic Target Recognition
Defence Science Journal · 27.10.2026 · 76 (6) · 914-922
DOI: 10.14429/dsj.22809 ↗Kurz gefasst
Fachartikel in der Zeitschrift Defence Science Journal, online erschienen am 27.10.2026. Verfasst von Vishwakarma Rahul Lalmani und 3 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 2.5).
Abstract (Originalsprache)
Synthetic Aperture Radar (SAR) automatic target recognition for defence applications requires models that are accurate, compact, and computationally efficient. This paper proposes EffiSARNet, a lightweight gated multi-stream network for SAR target classification. The architecture combines a shallow convolutional stem with hierarchical Doubly-Stochastic Gated Multi-Stream (DSGM) blocks to enable controlled inter-stream interaction, depthwise spatial filtering, contextual feature gating, and residual information flow. This design preserves target-discriminative scattering cues while maintaining a small computational footprint. Experiments on the ten-class MSTAR benchmark were conducted using validation-based checkpoint selection and independent test-set evaluation. EffiSARNet achieved 99.67 % test accuracy and a 99.67 % weighted F1-score while using only 29,868 trainable parameters, requiring 18.855 million MACs and 37.710 million FLOPs, with an estimated single-precision storage footprint of 0.123 MB for a 128 × 128 input chip. Under the same evaluation protocol, EffiSARNet outperformed several lightweight and deeper convolutional baselines while requiring substantially fewer parameters. These results indicate that EffiSARNet is well suited for deployment-oriented SAR target recognition in resource-constrained environments.
Zitieren
APA
Lalmani, V. R., Singh, R. S., Doddamani, G. R., & J., S. R. (2026). EffiSARNet: A Lightweight Gated Multi-Stream Network for Efficient SAR Automatic Target Recognition. Defence Science Journal, 76(6), 914-922. https://doi.org/10.14429/dsj.22809
Vancouver
Lalmani VR, Singh RS, Doddamani GR, J. SR. EffiSARNet: A Lightweight Gated Multi-Stream Network for Efficient SAR Automatic Target Recognition. Defence Science Journal. 2026;76(6):914-922. doi:10.14429/dsj.22809
BibTeX
@article{Lalmani2026,
author = {Lalmani, Vishwakarma Rahul and Singh, Ravi Shankar and Doddamani, Govind Rao and J., Senthil Rengarajan},
title = {EffiSARNet: A Lightweight Gated Multi-Stream Network for Efficient SAR Automatic Target Recognition},
journal = {Defence Science Journal},
year = {2026},
volume = {76},
number = {6},
pages = {914--922},
doi = {10.14429/dsj.22809}
}
EUVIMED