关键词: 船舶目标分割, DeepLabV3+, 轻量化网络, 注意力机制, 迁移学习

Abstract: Ship target segmentation plays a crucial role in enhancing the automation level of marine monitoring systems. However, existing segmentation methods often face issues such as mis-segmentation, susceptibility to noise interference, and low processing efficiency, making them difficult to adapt flexibly and efficiently to real-world scenarios. To address these challenges in maritime ship target segmentation tasks, a segmentation algorithm is proposed based on an improved DeepLabV3+ network, termed MET-DeepLabV3+. During the image preprocessing stage, bilateral filtering, multi-scale Retinex algorithm, and linear transformation are employed to mitigate the effects of noise, weather, and other adverse factors. In terms of model design, the lightweight MobileNetV2 is used as the backbone network to reduce model complexity, and the ECA-Net attention mechanism is introduced to enhance the model's ability to capture multi-scale features. Additionally, a transfer learning approach is adopted, where feature weights pre-trained on the SeaShips dataset are applied during model training to further optimize segmentation performance. Experimental results demonstrate that the improved algorithm achieves an average interaction-over-union (IoU) of 91.62% and an average pixel accuracy (PA) of 92.87%, representing improvements of 2.13% and 1.11% over the baseline DeepLabV3+ model, respectively. Moreover, it outperforms segmentation models such as HRNet, PSPNet, and UNet, which effectively meets practical demands of ship target segmentation tasks and offering significant application value.

Key words: ship target segmentation, DeepLabV3+, lightweight network, attention mechanism, transfer