【作者】 张习文；

【导师】 欧宗瑛；

【作者基本信息】 大连理工大学 ， 机械制造及其自动化， 2000， 博士

【摘要】 扫描工程图样识别是图象处理、模式识别和人工智能等多种学科的综合应用，直接面向企业需求，具有很高的理论意义和应用价值，是CAD领域的重要课题。经过多年研究，扫描工程图样识别已经取得较大进展，部分实现象素到矢量的转换。但是，现有识别方法多拘泥局部，串行处理，已实现的识别能力与质量离实际需求还有很大差距，识别理论和方法有待突破。 工程图样是工程图元的有机集合，可看作线条图形，而扫描工程图样是象素的自然集合。为将象素聚合为工程图元，本文力求加强表达单元的整体性，提高表达层次，重视各种关联，根据启发信息选择和组织识别数据和知识，进行分层次处理和自组织推理。本文提出一种称为单义域的新的图象表达单元，先将象素矩阵转化为单义域及其拓扑关系的集合。同一层次数据之间相互关联。不同层次数据之间也相互关联。识别是一个反复自组织的推理过程。本文提出基于单义域邻接图的扫描工程图样自组织智能识别。在算法实现上，采用模糊分类、遗传算法和面向对象知识表示等。处理过程分为三个阶段： （1）提取扫描工程图样的线条特征，构建单义域邻接图来表达形状 与拓扑信息，统一表达字符和图形的结构特征，使后续的识别 处理能在较高层次的基本单元上进行。 （2）遍历单义域邻接图，提取字符及其笔划特征，提取完整的几何 图元，采用矢量邻接图来组织获取的信息。 （3）基于矢量邻接图，采用面向对象知识表示来组织矢量之间的约 束知识，组合工程图元，同时提取关联信息，从而构建工程图 元邻接图。 在识别单义域、矢量和工程图元时，既注重横向的关联又重视纵向的关联。横向关联指图面不同部分之间的直接关联，纵向关联指单义域和它所构成的高层工程图元之间的关联。在识别中，先获得部分特征再生成整体结构，然后又用整体结构去指导部分特征的进一步把握；高层信息从低层数据获得，反过来又去指导低层数据，不同层次之间相互作用。在智能推理中，根据启发信息自动选择相应的识别知识，同时不断调整识别参数，以适应不断的数据变化。自组织识别在关联数据相互依赖、相互影响和相互作用的协作中进行，错综复杂的关系相互协调完成整个识别。 工程图样包含的图形和字符均可看作线条，线条之间存在多种连接关系。扫描工程图样识别先要获取图象的线条及其关系表达。本文所提出的单义域表达单元扩大了连通域表达范围，包括线段、圆弧、箭头和交点。采用游程邻接图表达二值图象，然后作深度优先遍历，基于游程宽度和拓扑一致形成条形域。引人模糊逻辑对条形域进行分类，获取初步矢量信息，对其中多义域做单义分裂，以线段和圆弧为基元，采用遗传算法来实现。单义域是具有矢量特征的局部象素合理聚集，反映关联象素的整体特性。继承游程的拓扑关系，构建单义域邻接图。基于单义域邻接图自组织识别扫描工程图样，在处理效率和抑制噪音误差影响等方面更为优越。 在单义域邻接图基础上，可对字符和图形进行自组织识别，在字符笔划域基础上进行字符提取，同时提取其笔划特征，为将来字符识别提供结构信息。根据字符域大小对字符域外接矩形进行自适应膨胀。根据字符域膨胀矩形相交来判定字符邻近程度，再加上字符共线为判据来生成字符串域。利用同串字符的外接矩形中心和所附图形对字符进行定向。在提取线“段、圆弧和圆时，先从种子域线段或圆弧出发，按照同线或同圆的要求识别处理，进行邻接图深度捏索，种子矢量不断生长，几何参数不断调整，从而获得完整信息。提取的信息采用矢量邻接图来组织。 上述工作可以实现多种扫描工程图样中的字符、线段、圆弧、圆和箭头的提取。但是，工程图样是工程图元的集合。本文采用面向对象方法，对工程图元进行对象设计。给出基于矢量邻接图的工程图元识别方法，根据已经提取的几何图元和字符信息，从某一特征图元出发，选择相应的识别知识（工程图元组成语法），搜索所有其它组元，提取完整信息。文中分别提取点划线和虚线的线段、圆弧和圆，还提取剖面线和尺寸，同时与约束图元关联。 上述识别方法已在开发的扫描工程图样识别原型系统中实现，软件采 一用面向对象和过程技术分析和设计。基于单义域邻接图的自组织识别方法丰富了扫描工程图样识别方法，加强宏观和整体处理能力，利用了更多的关联，力求对识别数据和知识进行自组织。对多种扫描工程图样进行识别，效果较好。更多还原

【Abstract】 Recognition of scanned engineering drawings is a comprehensive application that relates to multi-discipline, such as image processing, pattern recognition, and artificial intelligence, etc. The research on recognition is important in theory and practical application. The recognition is one of key issues in the field of CAD. Many progress in the area have been made, raster images can be partly transformed into vector data that can be used in a CAD system. However, many current approaches are limited on capturing local features and process sequentially, and the results recognized are not satisfactory, new approaches of recognition still have to be researched. An engineering drawing is a set of associated entities, which are line-like, and a scanned engineering drawing is composed of pixels. To extract entities from pixels, recognition process should be self-organizationally implemented by levels. The approach developed by author aims to capture more global features and intelligently infer with combining lower and higher local infonnation. The recognition approach places more emphasis on association relations among features. The data and knowledge are chosen and organized according to contexts of data. The recognition is to transform pixels into primitive regions. The features in the same level are associated each other. The features in different levels are also hierarchically associated each other. The recognition is to?infer self- organizationally using those associations. Scanned engineering drawings are self- organizationally recognized based on Primitive Region Adjacency Graph (PRAG). Author uses fuzzy classification, genetic algorithm and object-oriented knowledge representing in developing algorithms. The steps of processing is as follows: (1) To extract stripe features from an image, a PRAG is used to represent geometrical and topological data. It represents structure features of characters and graphics, and provides primary data for later processing. (2) Characters and their strokes, and integrated graphic primitives are extracted form the PRAG. Then, vectors recognized are stored in a vector adjacency graph (VAG). (3) Based on the VAG, constraint knowledge among vectors is organized using object-oriented knowledge representing. Entities are extracted, and associations among them are also extracted. The drawings are represented using an entity adjacency graph (EAG). Primitive regions, vectors and entities are extracted using association of features. In the same level, the global structure is constructed with local features Ill extracted. Then, it is to guild to capture local features. The recognition uses association of different levels. Knowlcdge of recognition is chosen automatically according to the start data, and some thresholds are adjusted during inferring to adapt to context changing. The associations depend on and affect each other. The association relations should be self-organized during the recognition implementation. Components of graphics and texts in drawings can be viewed as stripe regions. Recognition of images is to get stripes and their relationships. A new primitive region structure and a adjacency graph are developed for representing a scanned drawing. A primitive region can represent a line, an arc, an arrow or an intersection block, which enlarge the scope of quadrangle-like regions. A binary image is represented using a mn-length adjacency g更多还原