大型高温结构件表面缺陷三维动态检测理论与识别方法研究

The surface quality detection is urgent needed in the manufacturing process of large high temperature structure components. Non-contact vision detection is more suitable for the online real-time detection of the high temperature structural automated production line. However, for existing vision detection methods, the detection and recognition rate are still low caused by interference information. Moreover, the 3D depth information of the defect has not been detected. In order to solve the above problems, this project will employ the structure grating, stereo vision principle and deep learning theory to reconstruct the 3D morphology characteristics of surface defects. Meanwhile, a set of both temporal resolution, spatial resolution and accuracy of larger-scale high temperature structural components 3D surface defect detection and recognition system is build. The project focus on the multi-source stereo vision 3D reconstruction, suspected object detection and recognition for 2D surface defect image, surface defect discrimination and synchronous display fusing 3D point cloud and 2D image. At the same time, the method of theory research and experiment validation synchronous implementation, improve and perfect the related theory and complete system software development. This project not only has an important theoretical and practical significance to the 3D surface defect inspection of the larger-scale high temperature structural components with a better understanding, but also has a special significance for other defects detection system. Furthermore, this project will enrich and develop the traditional theory of the defect detection and recognition.

面向大型高温结构件制造过程中对表面质量检测的迫切需求，非接触式的视觉检测更适用于高温结构件自动化生产线上的在线实时检测。然而现有的视觉检测方法仍然存在干扰信息导致的检出率和识别率较低，以及未检测缺陷的三维深度信息等问题。针对上述问题，本项目以结构光栅、立体视觉原理和深度学习理论为基础，实时检测表面缺陷的三维形貌特征，进而构建一套同时兼顾时间空间分辨率和检测精度的大型高温结构件表面缺陷三维检测与识别系统。本项目重点开展多源立体视觉三维重构、二维表面缺陷图像可疑目标的检测与识别、以及三维点云与二维图像数据融合的表面缺陷判别与同步显示等方面的研究，同时将理论方法研究与实验验证同步实施，改进和完善相关理论方法，并完成系统软件的开发。本项目的研究不仅对大型高温结构件表面缺陷三维检测的深入理解具有重要的理论与实际意义，而且对其他缺陷检测系统的改进具有特别的促进意义，将丰富和拓展传统的缺陷检测与识别理论

大型金属结构件作为我国装备制造领域的关键部件，已广泛应用于高速铁路重轨、航空航天军用钢板等领域。对这些产品在生产过程中表面缺陷的实时监测，可及时发现缺陷并调控生产装备参数，进而提高产品的成材率。然而，基于机器视觉的高温态金属结构件表面缺陷实时检测识别系统，还存在时间空间分辨率与重构速度和精度的兼顾、小样本表面缺陷数据集与识别精度的协同、多维图像数据的深度融合等关键科学问题。针对以上关键问题，本项目深入开展了多源立体视觉三维重构、深度学习理论的表面缺陷检测与识别、小样本表面缺陷的图像分析方法、多维图像数据的深度融合检测等关键理论和技术的研究。实现了高温金属结构件表面缺陷三维高度、二维面积及缺陷类别等信息的综合检测与识别，突破了现有检测方法的局限性，形成全面、快速、高精度的表面缺陷检测方法，为我国高温金属结构件产品质量的监测与控制提供了重要的理论和技术支撑。此外，还开辟了工业界跨模态图像融合检测识别领域的一个新主题，为更广泛的工业应用提供了重要的理论参考。取得的成果不仅具有重要的科学意义，而且具有明确的工程应用背景和实际意义，同时对其他相关机器视觉检测识别的应用研究领域，也具有重要的启发和借鉴意义。