YOLO11s-EER: a lightweight small target detection algorithm for ship detection in remote sensing imagery

Detecting vessels plays a vital role in various domains, which include the safety of marine transportation, rescue operations, and the administration of harbor activities. However, current approaches still present challenges regarding real-time detection and lightweight deployment against complex backgrounds. This paper proposes a lightweight small ship detection model YOLO11s-EER for complex maritime remote sensing scenarios, aiming to improve the precision and efficiency of small vessel detection in complex backgrounds by utilizing remote sensing images. The proposed method leverages Extended Window Multi-Head Self-Attention (EW-MHSA) to establish global context awareness, physically suppressing ship-like noise such as wave glint and cloud shadow. By employing feature grouping without channel reduction and cross-spatial learning strategies, it preserves the edge and texture signals of tiny ships, addressing the issue of small target feature attenuation in deep networks. Furthermore, reparameterized convolution is adopted to achieve architecture decoupling characterized by "complex training and simple inference". The experimental findings demonstrate that YOLO11s-EER achieves mAP50 of 85.3% and 77.1% on the MASATI-v2 and LEVIR-ship datasets, respectively, with a 23.6% reduction in parameter count and a 58.2% decrease in computational load. Compared with existing methods, the proposed model achieves a better balance between accuracy and efficiency in complex maritime remote sensing scenarios, demonstrating strong generalization capability and potential for edge deployment.