Scholarlink

<p>Side-scan sonar imaging technology is vital to produce clearer imaging of ocean beds in low visibility conditions. However, sonar images suffer from low resolution, edge distortion, class imbalance and limited dataset for underwater target detection, and segmentation approaches. Addressing these challenges, we propose the Sonar Edge attention UNet a novel model that integrates a sobel operator based attention gate within UNet to enhance the edge feature representation of targets in the sonar images. With the proposed approach, underwater sonar target segmentation can be accurate in weak semantic information and complex background. To train and evaluate our model, we utilise a side-scan sonar images of AI4shipwrecks dataset and KLSG dataset. The study reveals proposed architecture able to capture edges in complex underwater environments having varying textures and shapes and mitigates the impact of target class imbalance. It achieves the highest meanIoU of 0.7647 and dice score of 0.6983 compared to the baseline UNet and Deeplabv3 models. The findings highlight the importance of custom-designed architectures tailored to the unique characteristics of sonar images, paving the way for advancements in the field of underwater sonar image segmentation and enhancing the visualization of submerged targets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SEAUNet: a novel edge-adaptive attention based UNet for sonar image segmentation

  • 期刊论文

  • 发表日期：

    2025年5月28日

查看全文
  • Divyabarathi G,
  • Gayathri Soman,
  • P. V. Sabeen Govind,
  • M. V. Judy

摘要

Side-scan sonar imaging technology is vital to produce clearer imaging of ocean beds in low visibility conditions. However, sonar images suffer from low resolution, edge distortion, class imbalance and limited dataset for underwater target detection, and segmentation approaches. Addressing these challenges, we propose the Sonar Edge attention UNet a novel model that integrates a sobel operator based attention gate within UNet to enhance the edge feature representation of targets in the sonar images. With the proposed approach, underwater sonar target segmentation can be accurate in weak semantic information and complex background. To train and evaluate our model, we utilise a side-scan sonar images of AI4shipwrecks dataset and KLSG dataset. The study reveals proposed architecture able to capture edges in complex underwater environments having varying textures and shapes and mitigates the impact of target class imbalance. It achieves the highest meanIoU of 0.7647 and dice score of 0.6983 compared to the baseline UNet and Deeplabv3 models. The findings highlight the importance of custom-designed architectures tailored to the unique characteristics of sonar images, paving the way for advancements in the field of underwater sonar image segmentation and enhancing the visualization of submerged targets.