EL-YOLO: an efficient and light YOLO algorithm for small object SAR ship detection

Synthetic aperture radar (SAR) image ship detection is a vital tool for maritime monitoring, illegal fishing, and traffic management due to SAR’s ability to operate in all weather and lighting conditions. However, there are still limitations in current SAR ship detection, resulting in a high false alarm rate of small targets due to speckle noise and complex near-shore interference caused by the imaging principle. Meanwhile, existing methods do not balance the detection of small targets and model weighting. To overcome these problems, we propose an efficient and lightweight SAR ship target detection method based on YOLOv8n, named efficient and light you only look once (EL-YOLO). This model comprises three essential components: small target extraction backbone (STEB), ultra-effective feature fusion neck (UE-Neck), and ultra-lightweight detection head (UL-Head). STEB is the backbone feature extraction network, significantly enhancing small and multiscale target extraction. UE-Neck can filter the redundant information and improve the information fusion of the small target ships to prevent the dilution of the small ship information. UL-Head provides a lightweight encoding and decoding format while ensuring detection accuracy. Extensive experiments evidence that EL-YOLO achieves an average precision of 98.0% on the SSDD and 92.2% on the high-resolution synthetic aperture radar image dataset (HRSID), the parameter is only 1.3M, and the computation is 7.8 Giga floating-point operations per seconds (GFLOPs) compared with the baseline, which is a 12.4% reduction in computation.