① 类别覆盖——九类标注是否全覆盖?材料、一般公差这两类"高幻觉"类别是否单独验证过[4]; ② 指标透明——能否提供分类别的精确率/召回/幻觉率,而不是一个总分?行业目前连统一多格式基准都还缺位[1][6]; ③ 输入兼容——矢量PDF、DWG之外,对扫描光栅件的抗噪表现如何[9]; ④ 下游打通——能否导出Excel/JSON等结构化清单并对接CMM、SPC系统[8]; ⑤ 人机协同——是否内置置信度、来源追溯与人工复核工作流,歧义案例能否自动转人工[2]; ⑥ 成本结构——领域适配的门槛有多高?研究表明LoRA式轻量微调以约1/10的标注成本即可达到全量微调85%~95%的性能,说明"小步快跑"式的试点完全可行[1];对中小企业,商用气泡图工具(如MBDVidia_BALLOON2D)与开源检测模型(如YOLOv11路线)的组合,往往比全自研更务实[10]。
Pagidi K R, Gardas A. Vision foundation models for engineering document intelligence and manufacturing inspection: a survey. Artificial Intelligence Review, 2026(预印本: Research Square, DOI: 10.21203/rs.3.rs-9528170/v1).
Khan M T, Chen L, Feng W, et al. Context-aware mapping of 2D drawing annotations to 3D CAD features using LLM-assisted reasoning for manufacturing automation. arXiv:2602.18296, 2026.
Khan M T, Chen L, Yong Z, et al. From drawings to decisions: a hybrid vision-language framework for parsing 2D engineering drawings into structured manufacturing knowledge. Robotics and Computer-Integrated Manufacturing, 2026, 99: 103186.
Khan M T, Yong Z, Chen L, et al. Automated parsing of engineering drawings for structured information extraction using a fine-tuned document understanding transformer. arXiv:2505.01530v3, 2025(IEEE IEEM).
Jong W R, Lin Y H, Lin Y C. Deep learning-based automation for converting 2D engineering drawings into 3D solid models. Machines, 2026, 14(7): 781.
Kumar A, Motiyani I, Kasturi T, et al. Enginuity: a dataset and benchmark for vision-language understanding of engineering diagrams. arXiv:2606.03410, 2026.
Kou Q, Shi X, Li Y, et al. MechVQA: benchmarking and enhancing multimodal LLMs on comprehensive mechanical drawing understanding. arXiv:2605.30794, 2026(ICML 2026).
Kim M, Kim Y, Kim H. Drawing-Recode: annotation grounding for parametric CAD code generation from raster 2D CAD drawings. arXiv:2607.27558, 2026.
Reddy T N, Kumar N, Ponnappa N P, et al. Intelligent GD&T symbol detection in mechanical drawings: a comparative study of YOLOv11, Faster R-CNN, and RetinaNet for quality assurance. Journal of Intelligent Manufacturing, 2026, 37(7): 2941-2959.