基于条件随机场的工程图轴类零件特征识别

Feature Recognition of Shaft Parts from Engineering Drawings Based on Conditional Random Fields

摘要: 工程图目前还是描述产品和指导生产的主要文件，从工程图中直接识别设计或制造特征可以充分利用现有技术资源，提高生产效率。但工程图隐含大量语义信息，很难被计算机提取和理解。条件随机场模型是一种可以自动整合各种特征、基于学习的分类技术，不依赖于形式化的启发式规则。采用条件随机场模型，提出了面向工程图的轴类零件特征识别方法。通过对轴特征在工程图中形状分析，定义和构建特征轮廓环及其关系，分析出特征轮廓环的属性特征和关系特征，构建出条件随机场无向图模型，通过对工程图样本的手工标记和模型训练，使训练模型的分类预测与实际需要的特征逐步一致，实现对轴工程图的特征识别。

Abstract: At present, the engineering drawing is the main document to describe products and guide their production. The direct recognition of design or manufacturing features from the engineering drawing can make full use of the existed technical resources and improve the production efficiency. However, engineering drawings contain a lot of semantic information, which is difficult to be extracted and understood by computers. The Conditional random field model is a kind of learning-based classification technologies that can automatically integrate all kinds of features without relying on formalizing heuristic rules. Based on conditional random field models, a feature recognition method from engineering draws of shaft parts is proposed. By analyzing the feature shape of shaft parts in the engineering drawing, the feature loops formed by contours of features and their relationships are defined. A set of attribute features and relation features of the feature loops are analyzed, by which the undirected graph of conditional random field models is constructed. By manual labeling for the engineering drawings and model training, the classification prediction of the trained model and the actually needed features are gradually consistent, which can realize the feature recognition of shaft parts from engineering drawings.