基于逆问题的机器视觉跨尺度测量理论建模及其应用技术研究

Measuring and detecting geometric features across scale from 10^-4mm to 10^2mm is an important but difficult task. In this project, we study the basic theory of machine vision cross-scale measurement and its application technology based on the inverse problem thought. First, we will set up an approach to describe the machine vision imaging scale space applying the state vector and search for the matching relation between imaging scale space and the multi-scale structure system of the imaged surface. We will also explore the connecting mechanism among scale subspaces and build the state transforming model of scale subspaces. Then, through investigating the bidirectional reflectance distribution function of each subspace and the irradiance equation of image, we will look for the texture correlation function under the micro-scale imaging condition and build the subspace imaging model that is based on the inverse problem theory. Third, We will research the superposition effect that the imaging conditions are mapped to the final images and develop the operators and algorithms which can be used to restrain or eliminate the facts that affect the accuracy of image measuring. The methods and algorithms applied to recover the real characteristics of measured objectives from the images of Non-Lambertian surfaces will be studied as well. Finally, integrating the scale subspace state transformation model and subspace imaging model, we will develop a state space model of machine vision cross-scale measurement based on the inverse problem. With the support of this model, the accuracy generating mechanism and the error propagating rule of cross-scale measuring will be investigated. In the process of project research, we will build an imaging experimental system with multi stations to examine the developed theory and models as well. This project has important theoretical significance and application value to the high precision automatic measurement in micro manufacturing.

针对10^-4mm–10^2mm范围跨尺度特征的测量与检测问题，本课题基于逆问题思想，开展机器视觉跨尺度测量基础理论与应用技术研究。建立机器视觉成像尺度空间的状态向量描述方法，寻求成像尺度空间与成像表面多尺度结构体系的匹配关系，探索尺度子空间的衔接机制，构建尺度空间状态变换矩阵；研究各子空间的双向反射分布函数和图像辐照方程，探索微尺度成像的纹理相关函数，建立基于逆问题的子空间成像模型；研究成像条件在最终图像中的叠加效应，开发抑制或消除影响图像测量准确性的算子与算法，研究从非朗伯表面所成图像中恢复被测目标真实特征的方法及算法；集成尺度空间状态变换模型和子空间成像模型，建立基于逆问题理论的机器视觉跨尺度测量状态空间模型，研究跨尺度测量的精度生成机制及误差传递规律；建立多工位成像实验系统，检验本课题所建立的理论及模型。该项目对微制造领域的高精度全自动测量具有重要的理论意义和应用价值。

项目针对10^-4mm–10^2mm范围跨尺度特征的测量与检测问题，基于逆问题思想，开展机器视觉跨尺度测量基础理论与应用技术研究。构建了基于逆问题的成像过程可控的机器视觉实验系统，以被测特征尺寸范围10^-2mm–140mm的航天密封圈为对象，建立了典型的跨尺度全向曲面零件的光学成像模型及其跨尺度智能测量与检测集成方法。提出基于逆问题的航天密封圈表面缺陷检测方法，建立了全向曲面零件的计算成像方法以及基于Ward反射模型的航天密封圈表面缺陷检测方法。将被测尺寸精度范围延伸至10^-4mm级别，建立了孔轴类零件跨尺度高精度智能测量方法。将被测尺度范围扩展至数十米量级，提出大尺度空间动态目标三维姿态角的视觉图像测量方法。针对弹目接近过程中的变尺度测量问题，提出动态变尺度红外目标的视觉测量与跟踪方法。将研究成果应用于智能装配线的研制，建立了基于智能测量的视觉在线标定方法及执行机构误差自动补偿方法。结合最新的深度学习方法，以计算资源受限的工业嵌入式平台应用为目标，提出基于深度学习的小样本航天密封圈表面缺陷检测算法，以及细粒度级表面缺陷轻量化检测算法。.通过上述研究，建立了集宏/微观测量于一体的机器视觉跨尺度自动测量理论，揭示机器视觉成像尺度空间内部及跨尺度空间之间复杂的多因素关系，掌握了非朗伯曲面微尺度成像过程的物理作用机理及跨尺度之间的衔接机制，探索并开发了基于深度学习的全向曲面类零件表面缺陷轻量化检测算法，掌握了机器视觉跨尺度测量的精度生成机制及误差传递规律，达到了智能制造科学对跨宏/微两个尺度自动测量与检测技术的要求。.应用项目成果，研制成功航天柔性零件智能测量与检测设备1台，以及武器系统舵机零件的精密选配测量设备1台，上述设备已在上海航天控制技术研究所得到应用，极大地提高了航天密封圈这类柔性曲面零件的尺寸测量精度与缺陷检测效率，解决了航空、航天工程以及重要密封系统中，柔性零件测检效率低、结果不稳定、耗费人力多的问题，有效提升了运载火箭和航天器密封系统的可靠性。