Scholarlink

<p>Automated crack detection in submerged concrete components of marine infrastructure—including offshore platforms, harbor facilities, bridge piers, and submarine tunnels—is critically hindered by complex optical degradation—including spectral attenuation, scattering-induced blur, and color distortion—which severely compromises deep learning-based perception systems. Although semantic segmentation offers a promising solution, fully supervised approaches are constrained by the extreme scarcity of pixel-level annotated underwater crack imagery. Furthermore, existing style transfer-based domain adaptation methods typically generate synthetic images that inadequately replicate authentic underwater degradation mechanisms. This paper presents a dual-stage semantic segmentation framework that integrates physics-aware style transfer with prior-enhanced semi-supervised learning. In the first stage, we augment CycleGAN with a composite physics-guided loss function that explicitly enforces fidelity to underwater imaging physics—constraining chromatic attenuation, scattering noise, and crack structural integrity—enabling high-fidelity translation from above-water to underwater crack imagery. In the second stage, we introduce a lightweight Physical Prior Module, embedding dual branches of dark channel prior and gray world assumption within a teacher—student semi-supervised network, facilitating physics-informed feature enhancement without additional annotations. Evaluations on a purpose-built multi-source underwater crack dataset demonstrate that our method achieves 63.78% mIoU on authentic underwater test images—a substantial gain of 6.26 percentage points over the baseline—while reducing Fréchet inception distance by 48.9% compared to the second-best style transfer approach. This work establishes a robust framework for automated structural integrity assessment of marine infrastructure in visually degraded underwater environments.</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

Two-Stage Semantic Segmentation of Underwater Concrete Cracks via Physics-Constrained Style Transfer and Prior-Augmented Semi-Supervised Learning

  • 期刊论文

  • 发表日期：

    2026年7月24日

查看全文
  • Shuai Teng,
  • Qinghua Xu,
  • Xiaosan Yin,
  • Shunli Xie,
  • Kaixin Chen,
  • Haodong Shi,
  • Mashiur Rahman

摘要

Automated crack detection in submerged concrete components of marine infrastructure—including offshore platforms, harbor facilities, bridge piers, and submarine tunnels—is critically hindered by complex optical degradation—including spectral attenuation, scattering-induced blur, and color distortion—which severely compromises deep learning-based perception systems. Although semantic segmentation offers a promising solution, fully supervised approaches are constrained by the extreme scarcity of pixel-level annotated underwater crack imagery. Furthermore, existing style transfer-based domain adaptation methods typically generate synthetic images that inadequately replicate authentic underwater degradation mechanisms. This paper presents a dual-stage semantic segmentation framework that integrates physics-aware style transfer with prior-enhanced semi-supervised learning. In the first stage, we augment CycleGAN with a composite physics-guided loss function that explicitly enforces fidelity to underwater imaging physics—constraining chromatic attenuation, scattering noise, and crack structural integrity—enabling high-fidelity translation from above-water to underwater crack imagery. In the second stage, we introduce a lightweight Physical Prior Module, embedding dual branches of dark channel prior and gray world assumption within a teacher—student semi-supervised network, facilitating physics-informed feature enhancement without additional annotations. Evaluations on a purpose-built multi-source underwater crack dataset demonstrate that our method achieves 63.78% mIoU on authentic underwater test images—a substantial gain of 6.26 percentage points over the baseline—while reducing Fréchet inception distance by 48.9% compared to the second-best style transfer approach. This work establishes a robust framework for automated structural integrity assessment of marine infrastructure in visually degraded underwater environments.