Scholarlink

<p>In this study, a deep learning-based classification approach was developed to identify corrosion categories from scanning electron microscopy (SEM) images of low-alloy medium-carbon steel. Steel specimens were exposed to HCl solutions under different conditions, and corrosion categories were assigned as C1, C2, C3, C4, C5, and CX based on mass loss measurements in accordance with ISO 9223. The problem was formulated as a multi-class image classification task, since each SEM image corresponds to a single corrosion category. A dataset consisting of 960 SEM image patches was constructed by dividing original SEM images into non-overlapping regions, thereby increasing sample diversity while preserving independent microstructural information. YOLO11s-cls and YOLO11m-cls models were trained using five-fold cross-validation with and without cosine learning rate scheduling. The best performance was achieved by YOLO11m-cls without cosine learning rate scheduling, yielding an accuracy of 0.736 ± 0.066 and a macro F1-score of 0.733 ± 0.067. The results demonstrate that SEM-based corrosion classification is feasible; however, misclassifications between morphologically similar corrosion levels indicate that the problem remains challenging, particularly for advanced corrosion stages.</p> Grpahical Abstract

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

Deep Learning-based Classification of Corrosion Categories from SEM Images of Low-Alloyed Medium Carbon Steel

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Ali Osman Gökcan,
  • Aysel Yazıcı,
  • M. Sadrettin Zeybek

摘要

In this study, a deep learning-based classification approach was developed to identify corrosion categories from scanning electron microscopy (SEM) images of low-alloy medium-carbon steel. Steel specimens were exposed to HCl solutions under different conditions, and corrosion categories were assigned as C1, C2, C3, C4, C5, and CX based on mass loss measurements in accordance with ISO 9223. The problem was formulated as a multi-class image classification task, since each SEM image corresponds to a single corrosion category. A dataset consisting of 960 SEM image patches was constructed by dividing original SEM images into non-overlapping regions, thereby increasing sample diversity while preserving independent microstructural information. YOLO11s-cls and YOLO11m-cls models were trained using five-fold cross-validation with and without cosine learning rate scheduling. The best performance was achieved by YOLO11m-cls without cosine learning rate scheduling, yielding an accuracy of 0.736 ± 0.066 and a macro F1-score of 0.733 ± 0.067. The results demonstrate that SEM-based corrosion classification is feasible; however, misclassifications between morphologically similar corrosion levels indicate that the problem remains challenging, particularly for advanced corrosion stages.

Grpahical Abstract