Scholarlink

<p>Central segregation is a typical internal defect in continuously cast slabs and can significantly affect the microstructural uniformity and final service performance of steel products. In current industrial practice, the evaluation of central segregation in acid-etched macrostructure images still largely depends on manual visual judgment, which suffers from subjectivity, poor consistency, and limited traceability. To improve the objectivity of macrostructure inspection, this study develops a pixel-level image analysis method for the digital extraction and quantitative characterization of central segregation in continuously cast slabs. An improved U-Net-based segmentation model, termed APRD-U-Net, was constructed to identify weak-boundary and irregular central segregation regions in acid-etched macrostructure images. The model integrates residual double convolution blocks, DropBlock regularization, an atrous spatial pyramid pooling module, and attention gates into the classical U-Net framework. These components enhance local feature representation, regularization capability, multi-scale context aggregation, and background-interference suppression, respectively. Experiments were conducted on an industrial acid-etched macrostructure image dataset, and the proposed model was compared with U-Net, Attention U-Net, and DeepLabV3+. The proposed APRD-U-Net achieved IoU, Dice coefficient, Precision, and Recall values of 0.919, 0.958, 0.947, and 0.969, respectively. Visual comparisons showed that APRD-U-Net produced more complete and continuous segmentation masks for blurred, low-contrast, and strip-like central segregation regions under complex acid-etched backgrounds. In addition, representative quantitative characterization demonstrated that the predicted masks could be converted into physically meaningful descriptors, including segregation area fraction, maximum length, average width, equivalent diameter, and largest-component area ratio. These results indicate that the proposed method has the potential to serve as a front-end segmentation and measurement tool for future grading-related digital assessment of central segregation in continuously cast slabs.</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

Pixel-Level Segmentation and Representative Quantitative Characterization of Central Segregation in Continuously Cast Slabs Based on Acid-Etched Macrostructure Images and an Improved U-Net Model

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Guoqing Du,
  • Sen Luo,
  • Weiling Wang,
  • Miaoyong Zhu

摘要

Central segregation is a typical internal defect in continuously cast slabs and can significantly affect the microstructural uniformity and final service performance of steel products. In current industrial practice, the evaluation of central segregation in acid-etched macrostructure images still largely depends on manual visual judgment, which suffers from subjectivity, poor consistency, and limited traceability. To improve the objectivity of macrostructure inspection, this study develops a pixel-level image analysis method for the digital extraction and quantitative characterization of central segregation in continuously cast slabs. An improved U-Net-based segmentation model, termed APRD-U-Net, was constructed to identify weak-boundary and irregular central segregation regions in acid-etched macrostructure images. The model integrates residual double convolution blocks, DropBlock regularization, an atrous spatial pyramid pooling module, and attention gates into the classical U-Net framework. These components enhance local feature representation, regularization capability, multi-scale context aggregation, and background-interference suppression, respectively. Experiments were conducted on an industrial acid-etched macrostructure image dataset, and the proposed model was compared with U-Net, Attention U-Net, and DeepLabV3+. The proposed APRD-U-Net achieved IoU, Dice coefficient, Precision, and Recall values of 0.919, 0.958, 0.947, and 0.969, respectively. Visual comparisons showed that APRD-U-Net produced more complete and continuous segmentation masks for blurred, low-contrast, and strip-like central segregation regions under complex acid-etched backgrounds. In addition, representative quantitative characterization demonstrated that the predicted masks could be converted into physically meaningful descriptors, including segregation area fraction, maximum length, average width, equivalent diameter, and largest-component area ratio. These results indicate that the proposed method has the potential to serve as a front-end segmentation and measurement tool for future grading-related digital assessment of central segregation in continuously cast slabs.