Scholarlink

<p>Accurate identification of object boundaries in images is essential for practical applications across science and engineering domains. This work presents a Wavelet-Enhanced Attention U-Net (WEA-U-Net), an improved segmentation architecture that integrates multi-scale wavelet decomposition and spatial attention modules within the U-Net framework. The model is developed to address segmentation challenges such as occlusions, varying illumination, and intricate structural details. Evaluated on publicly available plant leaf datasets, WEA-U-Net demonstrates notable improvements over established models including U-Net, SegNet, DeepLabV3, Mask R-CNN, and SE-HRNet, attaining a Dice coefficient of 0.93, Intersection-over-Union of 0.90, and pixel-level accuracy of 97%. These results indicate the method’s potential for broader use where precise delineation is required in complex images.</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

Enhancing fine-grained image segmentation: a wavelet-attention U-Net approach

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • Ishana Attri,
  • Lalit Kumar Awasthi,
  • Teek Parval Sharma

摘要

Accurate identification of object boundaries in images is essential for practical applications across science and engineering domains. This work presents a Wavelet-Enhanced Attention U-Net (WEA-U-Net), an improved segmentation architecture that integrates multi-scale wavelet decomposition and spatial attention modules within the U-Net framework. The model is developed to address segmentation challenges such as occlusions, varying illumination, and intricate structural details. Evaluated on publicly available plant leaf datasets, WEA-U-Net demonstrates notable improvements over established models including U-Net, SegNet, DeepLabV3, Mask R-CNN, and SE-HRNet, attaining a Dice coefficient of 0.93, Intersection-over-Union of 0.90, and pixel-level accuracy of 97%. These results indicate the method’s potential for broader use where precise delineation is required in complex images.