EfficientSARNet: A Lightweight, High-Performance, and Fast SAR Image Target Recognition Network

Synthetic aperture radar (SAR) image target recognition aims to identify objects captured in radar imagery accurately and efficiently. This core technology is widely used in various applications, including military and civilian contexts. However, SAR image target recognition models are typically deployed on edge devices with limited computational resources and real-time processing requirements, posing significant challenges to the model size, recognition accuracy, and inference speed. Existing SAR image target recognition methods struggle to achieve outstanding performance in all these aspects simultaneously. To address these challenges, we propose EfficientSARNet, a novel efficient model designed for SAR image target recognition. Specifically, we design a light-weight multi-branch (LWMB) feature extraction module to capture local diversity features efficiently. Additionally, we introduce an efficient linear self-attention module, proximal linear self-attention (PLSA), which leverages feature positional information for re-weighting to facilitate the capture of global features. EfficientSARNet is innovatively designed by both simply and efficiently stacking PLSA and LWMB modules in various stages to capture both global and local features, achieving light-weight, high performance, and fast SAR image target recognition under limited computational resources. Extensive experiments and thorough analysis on multiple subsets of the moving and stationary target acquisition and recognition (MSTAR) dataset and the OpenSARUrban dataset demonstrate that EfficientSARNet not only surpasses state-of-the-art algorithms in recognition accuracy but also shows advantages in metrics such as Parameters, MACs, and Throughput.