LSAN: A Novel Lightweight SAR Aircraft Target Detection Network

This paper proposes a novel lightweight SAR aircraft detection network (Lightweight SAR-Aircraft Network LSAN) based on YOLOv8n. In the backbone and neck part of the network, we design a lightweight feature extraction module with Efficient Multiscale Attention Module (EMA) to reduce the network parameters and computation cost. A multiscale Discrete semantic Feature Enhancement Module (DFEM) is added in the neck to enhance the capability of feature extraction. Experimental results on the China GF-3 SAR datasets demonstrate that the proposed algorithm achieves 98.1% mAP
⁵⁰
and 82.1% mAP
^50-95
at a speed of 122 FPS. Compared to the YOLOv8n model, LSAN shows an improvement of 1.4% in mAP
⁵⁰
and 4.8% in mAP
^50-95
. In particular, there is a reduction of 68.6% in the parameters and 30.8% of FLOPs. Also, experiments on other public SAR datasets demonstrate the generalization and effectiveness of our method.