基于视觉感知的图像语义建模及其在"智慧工厂"中的应用研究

"Smart factory" is the inevitable developing tendency of the factories in the future, which researches and develops the intelligent machine systems with self-awareness, self-learning and self-adaptive capacity similar to what human or creature have to improve the industrial production efficiency and the product quality. It will lead the transformation from traditional manufacturing to "green manufacturing". The visual monitoring in industrial processes is indispensable technical force in the "smart factory", but the intelligent perception of the current industrial vision system falls far short of the inherent demand of "smart factory". Based on the human visual perception characteristics, this project introduces such theories and methods including the statistical learning, pattern recognition and so on into the image semantic mining, which plans to research the optimal feature selection method for multi-vision and redundant images of complex target based on visual perception, the abstract semantic modeling and knowledge representation method with incremental learning ability for the multi-vision redundant images, as well as the machine perception of the internal semantics and intelligent reasoning of the environmental semantics of the industrial visual images; this project will bridge the "semantic gap" between low-level features and high-level semantic of images to the maximum extent and solve a series of problems that the "smart factory" oriented industrial vision system in the automatic detection and intelligent perception, which is difficult to automatically measure the visual target and lack in the effective evaluation for the accurate detection of complex target, and difficult to intelligently percept the external environment semantic. This project will consequently lay the theoretical foundation for the intelligent decision and automatic production in the "smart factory".

"智慧工厂"是未来工厂发展的必然趋势，通过研发具有人类或生物类似自我感知、自我学习和自我适应能力的智能机器系统来提高工业生产效率及产品质量，实现传统制造向"绿色制造"的转型。其中，工业过程视觉监控是必不可少的技术力量，但是当前工业视觉系统的智能感知能力还远达不到"智慧工厂"的内在需求。本项目从人类视觉的感知特性出发，将统计学习、模式识别等理论和方法引入到图像语义挖掘中，拟研究基于视觉感知的复杂目标多视觉图像最优特征选择方法、具有增量学习能力的多视觉冗余图像抽象语义建模与知识表达方法、工业视觉图像内部语义机器感知与环境语义智能推理方法；最大程度地弥合图像底层特征和高层语义之间的语义鸿沟，解决面向"智慧工厂"的工业视觉系统在自动检测与智能感知中存在的视觉目标自动测量难、缺乏目标检测准确性的有效评价手段、外部环境语义难以智能感知等一系列难题，为"智慧工厂"中的智能决策与自动化生产奠定理论基础。

“智慧工厂”是现代工厂信息化发展的新阶段，也是工业生产完全自动化之后的必然发展趋势。面向“智慧工厂”的工业视觉信息自动获取、分析与智能理解、生产过程的人机交互和智能决策等问题，是其中关键性研究课题。项目组对工业过程监测图像的统计分布特性进行了深入分析，根据生物视觉感知中早期生物视觉处理的主要步骤，将统计学习、模式识别等方法应用到了视觉图像语义建模与知识表达中，探讨了图像统计建模与贝叶斯推理在融合视觉监测图像信息的工业过程工况监测和产品质量检测中的可行性方案。提出了基于集成学习的交互式图像分割方法，该方法基于半监督学习中的聚类假设，提出一种REG-Boosting算法实现半监督分类器学习以实现工业监测目标的交互式半监督分割；针对工业监测图像前景目标和背景区域的分割结果，提出一种基于B样条概率密度函数估计的复杂图像目标空间结构统计建模方法，实现图像前景目标和背景区域空间结构特性的区别描述；对无明显前景目标的监测图像，提出一种基于双树复小波变换的复杂图像表面纹理特征表征方法，实现复杂监测图像信息的参量化描述；最后，通过融合工业过程监测中的视觉图像信息与过程工艺参量信息，提出一种基于多信息融合与可拓理论的工业过程工况识别方法。所提方法在锑矿浮选过程和某粮食加工企业的大米加工过程进行了初步应用，实现了工业过程工况状态和产品质量的智能感知，为面向“智慧工厂”中的工况智能化决策提供了有力支撑。