基于注意力特征融合的海底管线语义分割方法OA

As a critical component of marine infrastructure,the high-precision automatic segmentation of subsea pipelines is essential for safety monitoring and intelligent inspection.However,challenges such as complex seabed backgrounds,slender and faint pipeline targets,and extreme class imbalance severely limit the performance of existing segmentation methods.To address these challenges,this paper proposes an enhanced U-Net architecture.By introducing the Spatial-Channel Fusion Module(SDFM)and Context-Guided Attention Fusion Module(CGAF),it synergistically enhances the perception and recovery capabilities of pipeline structures.SDFM jointly models spatial and channel dependencies during early feature extraction,strengthening responses in critical regions.CGAF dynamically integrates jump-connected features through a local-global dual-path attention mechanism,precisely suppressing background interference and improving boundary accuracy.Experimental results demonstrate that the proposed method significantly outperforms mainstream models such as DeepLabv3+,U-Net,and the YOLO series,fully validating its superiority and robustness in segmenting slender pipeline objects within complex underwater scenarios..