MAMS-UNet: Hybrid Attention Network for Close-Range Ship Optical Image Segmentation

Ship target segmentation plays a critical role in marine monitoring systems. However, existing segmentation methods often encounter challenges such as missegmentation, noise interference, and low processing efficiency when dealing with ship targets in complex maritime environments. To address these issues, this paper proposes a segmentation algorithm based on the U-Net architecture, termed MAMS-UNet, specifically designed for maritime ship target segmentation tasks. Firstly, a novel Median-Mean-Max Channel Attention Module (MAMS) based on multi-scale spatial convolution is designed. By integrating global pooling and multi-scale depthwise convolutions, the module enhances the model’s ability to extract complex features. Secondly, the MAMS module is incorporated into the skip connections of U-Net, optimizing the feature transmission process and significantly improving segmentation accuracy for small targets and edge regions. Experimental results demonstrate that MAMS-UNet achieves a mean Intersection over Union (MIoU) of 90.32% and a mean Pixel Accuracy (MPA) of 95.07%, representing improvements of 1.24% and 0.58%, respectively, over the traditional U-Net model. Moreover, MAMS-UNet outperforms other state-of-the-art segmentation models such as PSPNet, HRNet, DeepLabV3+, and SegNet, effectively meeting the practical requirements of maritime ship target segmentation. These findings highlight its high application value and broad potential for deployment in real-world marine monitoring tasks.