Scholarlink

<p>Vision foundation models such as SAM, CLIP, DINOv2, and multimodal large language models (MLLMs) are now common starting points for general computer-vision tasks. Their use in engineering visual data, however, is still split across separate communities working on drawing understanding, document intelligence, and manufacturing inspection. We found no prior survey that studies engineering drawing understanding and manufacturing inspection together from a foundation-model perspective, although both domains face the same basic obstacle: models pretrained on natural images must be adapted to thin-line, symbol-dense engineering documents and to controlled-lighting, texture-repetitive inspection imagery. Using a PRISMA-informed protocol, this survey reviews 94 studies published from 2020 to 2026 across six databases. For engineering drawing understanding, we cover foundation-model methods for symbol detection, GD&amp;T extraction, table and BOM parsing, P&amp;ID graph construction, drawing-to-CAD reconstruction, and floor-plan analysis. For manufacturing inspection, we follow the progression from CLIP-based zero-shot anomaly detection to DINOv2 self-supervised methods and MLLM-based defect reasoning, separating benchmark performance from evidence of factory readiness. Across the two domains, we identify five SAM adaptation strategies, compare adaptation regimes from frozen prompting to LoRA fine-tuning, catalogue 16+ engineering-specific datasets, and analyse six benchmark gaps, including the lack of a standard multi-format engineering drawing benchmark. We present DrawingBench as a five-track benchmark <i>design proposal</i> (data sources, baseline lineup, and evaluation protocol for each track) and outline research directions for engineering-specific pretraining corpora, parameter-efficient adaptation libraries, edge-deployment pipelines compatible with manufacturing execution systems and programmable logic controllers, and multimodal engineering agents with metrologically traceable perception.</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

Vision foundation models for engineering document intelligence and manufacturing inspection: a survey

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月27日

查看全文
  • Kishore Reddy Pagidi,
  • Anusha Gardas

摘要

Vision foundation models such as SAM, CLIP, DINOv2, and multimodal large language models (MLLMs) are now common starting points for general computer-vision tasks. Their use in engineering visual data, however, is still split across separate communities working on drawing understanding, document intelligence, and manufacturing inspection. We found no prior survey that studies engineering drawing understanding and manufacturing inspection together from a foundation-model perspective, although both domains face the same basic obstacle: models pretrained on natural images must be adapted to thin-line, symbol-dense engineering documents and to controlled-lighting, texture-repetitive inspection imagery. Using a PRISMA-informed protocol, this survey reviews 94 studies published from 2020 to 2026 across six databases. For engineering drawing understanding, we cover foundation-model methods for symbol detection, GD&T extraction, table and BOM parsing, P&ID graph construction, drawing-to-CAD reconstruction, and floor-plan analysis. For manufacturing inspection, we follow the progression from CLIP-based zero-shot anomaly detection to DINOv2 self-supervised methods and MLLM-based defect reasoning, separating benchmark performance from evidence of factory readiness. Across the two domains, we identify five SAM adaptation strategies, compare adaptation regimes from frozen prompting to LoRA fine-tuning, catalogue 16+ engineering-specific datasets, and analyse six benchmark gaps, including the lack of a standard multi-format engineering drawing benchmark. We present DrawingBench as a five-track benchmark design proposal (data sources, baseline lineup, and evaluation protocol for each track) and outline research directions for engineering-specific pretraining corpora, parameter-efficient adaptation libraries, edge-deployment pipelines compatible with manufacturing execution systems and programmable logic controllers, and multimodal engineering agents with metrologically traceable perception.