Smart three-dimensional machine vision

• Main Author: Huang, Chongtian
• Other Authors: Anand Krishna Asundi
• Format: Theses and Dissertations
• Language: English
• Published: 2019
• Subjects: DRNTU::Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/102839; http://hdl.handle.net/10220/48588
• Institution: Nanyang Technological University

Three-dimensional inspection is increasingly being demanded by industry in various sectors and none more so than in wafer fabs. While there are many three-dimensional surface measuring instruments, there is still a gap to their adoption in the industry due to lack of methodologies for measuring large objects with micrometer axial resolution and at the same time to zoom into small regions with nanometer resolution. Apart from this, there is also a need to automate the inspection process. In this report, non-interferometric schemes are developed and improved for inspection of specular objects at different scales and a proposal for identifying defects with deep learning over this large field which can then be imaged in three dimensions using other novel systems with higher resolution. Firstly, for wafer flatness and warpage measurement, a Phase Measuring Reflectometry (PMR) with a large field of view is enhanced to measure wafers as large as 300 mm with no-scanning or moving components. Compared with existing scanning-based systems, this provides full-field, high axial resolution (few micrometers) measurement capabilities of wafers. Wafer warpage during various stages of processing can be measured rapidly thus providing a basis to modify or improve the processing parameters. The system which uses dual camera approach for measurement using structured light also allows for re-positioning and registration of the sample. Furthermore, We can obtain the rectified wafer image and global surface profile by the inspection system. Next, from the processed high-resolution images, defects can be highlighted. Traditionally, this would be done using image processing methods which are generally slow and not precise. To improve the robustness and reliability of defect detection, we introduce deep-learning methods for two-dimensional defects localization and segmentation. This approach requires training the machine to learn different defect shapes and sizes which then can automatically identify defects on test samples. Application of this can be more widespread, and we have demonstrated the scratch and dig measurement of optics as well as for defect detection in the electronics industry. Usually, the PMR system can resolve defects, but the axial resolution of a few micrometers is not sufficient to get accurate three-dimensional profiles. Furthermore, the wafer has many dies which need to be inspected at a higher resolution. Towards this end, We developed a novel multi-scale system by improving a commercial device built by a local company. This system can measure three-dimensional profiles in two ways: first, using the principle of the Transport of Intensity Equation (TIE), the system can obtain three-dimensional profiles of defects or dies with nanometer axial resolution. Furthermore, by attaching the system to an industrial microscope, high spatial resolution can be obtained as well. Secondly, if the height of objects is much larger compared to the wavelength of visible light, a deconvolution-based approach is used to extract the depth from a stack of images. In the current system, the focus scanning is in high speed with the assistant of a novel deep learning method thus improving the processing times over conventional confocal-type systems. Furthermore, the deconvolution-based depth retrieval method does not require the usage of a pinhole as used in confocal systems. With learning methods, the sample's shape can be reconstructed by computer without intensive iteration. Finally, an aluminum sample inspection results are shown to prove the performances of the proposed system scheme. Compared with conventional methods such as point or line scanning methods, this scheme can obtain the whole surface profile without scanning or contact, and with the deep learning algorithm and the multi-scale measurement scheme, a single inspection system can inspect not only the global surface profile but also the details of the three-dimensional structure of defects plus devices. This methodology can be adopted in many three-dimensional machine vision applications for inspecting digs and scratches of optical components as well as wavefront analysis, in MEMS and micro-systems arena for wafer warpage and die inspection in three-dimensional, as well as other consumer applications.