Scholarlink

<p>To address the critical challenges of micro-defect recognition, multi-scale feature loss, and edge blurring in underwater structural spalling segmentation, this paper proposes a multimodal fusion semantic segmentation network based on DeepLabV3+. The model employs MobileNetV2 as the backbone, integrated with a Squeeze-and-Excitation channel attention mechanism to enhance discriminative feature representation. In the ASPP module, standard atrous convolutions are replaced with depthwise separable convolutions to construct a DenseASPP architecture. Dense connections and a combined cross-entropy and Dice loss function are introduced to improve the extraction capability for underwater spalling defects. Furthermore, a Wavelet Transform is incorporated into the encoder to guide the defect feature extraction process, further refining the segmentation accuracy of spalling edges. Experimental results demonstrate the superior performance of the proposed method, achieving a Mean Accuracy (MA) of 0.96, Mean Boundary F1-score (MBFscore) of 0.80, and Mean Intersection over Union (MIoU) of 0.89. Compared to other state-of-the-art methods, the proposed approach improves MA by 4.52–17.42%, MBFscore by 6.77–66.12%, and MIoU by 3.61–22.49%, validating its effectiveness in segmenting surface spalling defects on underwater concrete structures and providing a new technical pathway for underwater defect recognition.</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

Precise Edge Segmentation of Underwater Structural Spalling Using an Improved DeepLabV3+ with Wavelet Transform and Dense Connections

  • 期刊论文

  • 发表日期：

    2026年2月13日

查看全文
  • Yixiao Zhang,
  • Airong Liu,
  • Shuai Teng,
  • Bingcong Chen,
  • Jialin Wang,
  • Zhihua Wu,
  • Jiawei Zhang,
  • Hailiang Chen

摘要

To address the critical challenges of micro-defect recognition, multi-scale feature loss, and edge blurring in underwater structural spalling segmentation, this paper proposes a multimodal fusion semantic segmentation network based on DeepLabV3+. The model employs MobileNetV2 as the backbone, integrated with a Squeeze-and-Excitation channel attention mechanism to enhance discriminative feature representation. In the ASPP module, standard atrous convolutions are replaced with depthwise separable convolutions to construct a DenseASPP architecture. Dense connections and a combined cross-entropy and Dice loss function are introduced to improve the extraction capability for underwater spalling defects. Furthermore, a Wavelet Transform is incorporated into the encoder to guide the defect feature extraction process, further refining the segmentation accuracy of spalling edges. Experimental results demonstrate the superior performance of the proposed method, achieving a Mean Accuracy (MA) of 0.96, Mean Boundary F1-score (MBFscore) of 0.80, and Mean Intersection over Union (MIoU) of 0.89. Compared to other state-of-the-art methods, the proposed approach improves MA by 4.52–17.42%, MBFscore by 6.77–66.12%, and MIoU by 3.61–22.49%, validating its effectiveness in segmenting surface spalling defects on underwater concrete structures and providing a new technical pathway for underwater defect recognition.