基于深度学习的柱大样构件CAD识别研究方法
Research Method for Detailed Recognition of Column Component Details Based on Deep Learning
为满足智能化工程生产需求，本研究提出几何感知的YOLO扩展模型GYOLO-U，用于CAD图纸中柱大样详图的自动识别和线段分类。该模型具备三大创新：(1) 双分支架构同步预测目标边界框与几何基元；(2) 基于局部块的最大角度检测器提取柱大样显著特征；(3) 两阶段训练策略，先对几何图案进行预训练，再基于2800张真实柱大样CAD数据微调。系统整体识别精度达89.11%，单图识别加后期推理分层约耗时4.5秒，较人工点选确认构件分层精度提升9.28%。
工程设计、工程测绘计算技术、计算机技术自动化技术、自动化技术设备
CAD识别YOLOv8几何深度学习局部角度检测两阶段训练
周剑峰,陆莉军,李奎.基于深度学习的柱大样构件CAD识别研究方法[EB/OL].(2025-09-24)[2025-09-29].http://www.paper.edu.cn/releasepaper/content/202509-28.点此复制
To meet the needs of intelligent engineering production, this study proposes a geometry-aware YOLO extension model named GYOLO-U for the automatic recognition and line segment classification of column detail drawings in CAD designs. The model features three key innovations: (1) a dual-branch architecture that simultaneously predicts target bounding boxes and geometric primitives; (2) a local patch-based maximum angle detector for extracting salient features of column details; (3) a two-stage training strategy involving pre-training on geometric patterns followed by fine-tuning on 2,800 real-world CAD column detail images. The system achieves an overall recognition accuracy of 89.11%, with a processing time of approximately 4.5 seconds per image for recognition and subsequent hierarchical inference, improving hierarchical annotation accuracy by 9.28% compared to manual component selection.
CAD recognitionYOLOv8geometric deep learningpatch-based angle detectiontwo-stage training
展开英文信息
评论