Scholarlink

<p>A novel image-based corrosion assessment method for low-alloy steels by integrating deep learning with quantitative metallography is proposed. The critical challenges of in situ corrosion monitoring in marine environments are addressed by introducing three key innovations: (1) a five-stage corrosion classification scheme based on combined macro- and micro-scale morphological analysis; (2) a high-quality annotated dataset (LASCD-UHD-2025) containing 1000 ultra-high-resolution images (5472 pixel × 3648 pixel) of Q420 and Q420RE steels under 3.5 wt.% NaCl immersion; and (3) an advanced UNet architecture enhanced with global attention module, triplet attention mechanism, and a Canny edge-weighted loss function. The advanced UNet achieves 53% mean intersection over union and 56% mean average precision, which is about 18% and 15% higher than the baseline UNet. Validation via gravimetric corrosion test, electrochemical characterization, and microscopic analysis confirms the reliability of the segmentation results. Using this method, we quantify that rare earth-alloyed Q420RE steel exhibits an approximately 14% slower corrosion progression compared to conventional Q420 steel during the initial corrosion stage. This integrated method enables automated, non-destructive corrosion monitoring, and offers broad potential for in situ real-time corrosion monitoring of materials used in marine environment.</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

Advanced UNet for high-resolution semantic segmentation of low-alloy steel corrosion evaluation and prediction

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Wen-Long Wang,
  • Jie Yang,
  • Li Tan,
  • Qing-Shuai Zhang,
  • Zong-Zhe Man,
  • Li-Wei Wang,
  • Jia-Jun Liu,
  • Zhong-Yu Cui

摘要

A novel image-based corrosion assessment method for low-alloy steels by integrating deep learning with quantitative metallography is proposed. The critical challenges of in situ corrosion monitoring in marine environments are addressed by introducing three key innovations: (1) a five-stage corrosion classification scheme based on combined macro- and micro-scale morphological analysis; (2) a high-quality annotated dataset (LASCD-UHD-2025) containing 1000 ultra-high-resolution images (5472 pixel × 3648 pixel) of Q420 and Q420RE steels under 3.5 wt.% NaCl immersion; and (3) an advanced UNet architecture enhanced with global attention module, triplet attention mechanism, and a Canny edge-weighted loss function. The advanced UNet achieves 53% mean intersection over union and 56% mean average precision, which is about 18% and 15% higher than the baseline UNet. Validation via gravimetric corrosion test, electrochemical characterization, and microscopic analysis confirms the reliability of the segmentation results. Using this method, we quantify that rare earth-alloyed Q420RE steel exhibits an approximately 14% slower corrosion progression compared to conventional Q420 steel during the initial corrosion stage. This integrated method enables automated, non-destructive corrosion monitoring, and offers broad potential for in situ real-time corrosion monitoring of materials used in marine environment.