Vision inspection as an important automation technology in manufacturing industry undertakes many appearance quality related tasks, such as product positioning, recognition, and measurement. Intelligent manufacturing increases the demand for vision methods in production scenes. However, the diversity and increment of the targets, and the feature robustness and scheme migration have restricted further implementation of this technology. To address this problem, this study establishes a robust inspection system, including type insensitive target ontology rough localization, component-based model reverse incremental recognition, and geometric attribute extraction, based on the joint strategy of lightweight neural network recommendation and knowledge guidance. Domain common sense and artificial knowledge are integrated into the whole method, of which the reasonable structural design alleviates the limitation of reliance on artificial intelligence or traditional image processing alone. Taking the connector in the electronic and electrical industry as an example, the proposed method demonstrates successful model recognition and component position extraction of 48 types (including two new models) of connectors, and significant adaptivity to the target posture, imaging environment, and feature diversity.

视觉检测作为制造业中一项重要的自动化技术，承担着许多与外观质量相关的任务，如产品定位、识别、测量等。智能制造增加了生产场景对视觉方法的需求。然而，目标的多样性和增量、特征鲁棒性和方案迁移限制了该技术的进一步实施。针对这一问题，本研究基于轻量级神经网络推荐和知识引导的联合策略，建立了一个鲁棒的检测系统，包括类型不敏感目标本体粗定位、基于组件的模型反向增量识别和几何属性提取。将领域常识和人工知识整合到整个方法中，其中合理的结构设计减轻了单独依赖人工智能或传统图像处理的局限性。以电子电气行业的连接器为例，该方法证明了48种连接器（包括两种新模型）的成功模型识别和元件位置提取，对目标姿态、成像环境和特征多样性具有显着的适应性.