Scholarlink

<p>In continuous casting, accurate prediction of mold flux viscosity is essential for maintaining quality of casting product. To overcome the bottleneck in balancing extraction efficiency, data purity, and physical fidelity, this study evaluates four distinct graphical data mining paradigms across a benchmark dataset of 300 publications on mold flux viscosity. These frameworks specifically include Method A based on system call binary encapsulation, Method B based on native stream parsing, Method C based on visual coordinate rendering and cropping, and Method D based on object-level stream parsing and heuristic filtering optimization. Results show that Methods A and B offer high recall (98.96&#xa0;pct), but they suffer from excessive redundancy and CMYK-induced color inversion (<i>D</i><sub>hist</sub>&#xa0;=&#xa0;1.1804). Method C avoids color inversion but exhibits severe resolution degradation (<i>R</i><sub>sfr</sub>&#xa0;&lt;&#xa0;0.1700) and structural fragmentation. In contrast, the proposed Method D achieves the optimal balance by leveraging optimized object-level stream parsing, reaching an efficiency of 0.193&#xa0;s/page, a 98.98&#xa0;pct recall, and a minimal invalid semantic image ratio of 4.90&#xa0;pct. By integrating ICC-based color mapping, Method D rectifies chromatic distortions with a <i>D</i><sub>hist</sub> of 0.0897 while preserving microstructural grain details with an <i>R</i><sub>sfr</sub> value ranging from 0.81 to 0.90. This methodology provides robust technical support for high-precision databases of mold flux viscosity, advancing the transition toward a data-intelligent research paradigm.</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

Multimodal Mining of Mold Flux Literature for Viscosity Analysis in Continuous Casting

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Yuanjin Mu,
  • Wanlin Wang,
  • Lejun Zhou,
  • Il Sohn,
  • Sibao Zeng,
  • Liwu Zhang

摘要

In continuous casting, accurate prediction of mold flux viscosity is essential for maintaining quality of casting product. To overcome the bottleneck in balancing extraction efficiency, data purity, and physical fidelity, this study evaluates four distinct graphical data mining paradigms across a benchmark dataset of 300 publications on mold flux viscosity. These frameworks specifically include Method A based on system call binary encapsulation, Method B based on native stream parsing, Method C based on visual coordinate rendering and cropping, and Method D based on object-level stream parsing and heuristic filtering optimization. Results show that Methods A and B offer high recall (98.96 pct), but they suffer from excessive redundancy and CMYK-induced color inversion (Dhist = 1.1804). Method C avoids color inversion but exhibits severe resolution degradation (Rsfr < 0.1700) and structural fragmentation. In contrast, the proposed Method D achieves the optimal balance by leveraging optimized object-level stream parsing, reaching an efficiency of 0.193 s/page, a 98.98 pct recall, and a minimal invalid semantic image ratio of 4.90 pct. By integrating ICC-based color mapping, Method D rectifies chromatic distortions with a Dhist of 0.0897 while preserving microstructural grain details with an Rsfr value ranging from 0.81 to 0.90. This methodology provides robust technical support for high-precision databases of mold flux viscosity, advancing the transition toward a data-intelligent research paradigm.