LW-SARDet: A Lightweight SAR Ship Detector via Decomposed Convolution

Synthetic aperture radar (SAR) has been widely utilized in ship detection due to its all-weather and all-day operational capabilities. Deep learning method, known for its exceptional detection precision, has emerged as the predominant approach for SAR ship detection. However, existing high-precision detectors are too complex to be deployed on airborne platforms. To address this challenge, based on the YOLOv8n architecture, this study proposes a lightweight SAR ship detector (LW-SARDet), which incorporates three lightweight designs. First, LW-SARDet decreases the downsampling factor of YOLOv8n by considering the small target characteristics in SAR ship scenes. Second, LW-SARDet introduces the efficient-neck, a lightweight feature fusion module that efficiently integrates high-level and low-level features by integrating channel attention mechanisms. At last, LW-SARDet proposes decomposed convolutions to replace standard convolutions, leading to further effective model compression. The experimental results indicate that LW-SARDet has achieved a state-of-the-art performance by utilizing only 0.4 M parameters and 2.9 GFLOPs. It has attained a 98.7% AP50 on the SAR ship detection dataset (SSDD).