MLANet: A Robust Ship Segmentation Network Based on Multilevel Multiattention Feature Fusion for Complex Maritime Background Environments

Abstract

Synthetic aperture radar (SAR) is a powerful sensor for long-range, all-weather, and large-scale surveillance, making SAR-based ship semantic segmentation a research hotspot. However, accurate segmentation ships, especially small vessels, in near-port waters remain a challenge due to complex background interference in SAR images. Current methods often struggle to extract sufficient features for small ships, leading to high missed detection rates. Furthermore, the complex oceanic background increases false detection rates. To address these issues, we propose MLANet, a multilevel feature-enhanced, multiattention fusion network specifically designed for ship segmentation in SAR images. MLANet leverages the strengths of both convolutional neural network (CNN) and Transformer to perform efficient multiscale feature extraction. The feature enhancement module (FEM) refines global and local features, retaining critical information for small ships, while the attention fusion module reduces background interference. Additionally, a hybrid loss function emphasizes both the shape and boundary of vessels during segmentation. The experimental results show that MLANet achieves 94.83% mean pixel accuracy (mPA) and 90.66% mean intersection over union (mIoU) on the SAR ship detection dataset (SSDD), and 91.74% mPA and 87.72% mIoU on the high-resolution SAR image dataset (HRSID), demonstrating its strong competitiveness and effectiveness in challenging environments.