| Author: | Gao, Sanshan |
| Title: | Deep learning-based resolution-enhanced autostereoscopic three-dimensional surface metrology |
| Advisors: | Cheung, C. F. Benny (ISE) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Subject: | Metrology Surfaces (Technology) -- Measurement Deep learning (Machine learning) Image processing Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Industrial and Systems Engineering |
| Pages: | xxv, 181 pages : color illustrations |
| Language: | English |
| Abstract: | Precision measurement for micro-structured surfaces is important for the fabrication of micropatterns to guarantee form accuracy. Contact and non-contact measurement methods alike have been extensively used to inspect these surfaces, where the optical sensor used in the non-contact methods does not cause probable damage to the measured parts, and the speed of data acquisition is much faster. Among these non-contact measurement technologies, an autostereoscopic measurement system with a rapid data acquisition process is an effective method to achieve accurate on-machine measurement. Autostereoscopy technology can provide a rapid and accurate three-dimensional (3D) measurement solution for micro-structured surfaces. The autostereoscopic 3D measuring system can record elemental images within one snapshot and the measurement accuracy can be quantified from the disparities existing in the 3D raw information. One of the primary challenges in improving the measurement resolution of autostereoscopic 3D systems is the natural compromise between spatial resolution resulting in finer details and angular resolution enabling accurate parallax restoration. This trade-off poses an obstacle to enhancing the resolution of the system. Improving angular resolution is of utmost importance when considering the trade-off of light field data. Within the domain of enhancing angular resolution through deep learning methods, advancements have been made in both non-depth-based techniques and the techniques requiring depth estimation. However, non-depth-based methods usually generate image ghosting when the light field images have large disparity ranges. It is difficult to realize accurate disparity estimation directly obtained through light field images for the depth-based methods, so image artefacts are usually fabricated in the novel views especially when scenes are complex. In addition, a lack of ground truth of the generated novel views makes the training on the finite data insufficient since the training data have to be split into inputs and their labels. In this thesis, a novel semi-supervised learning paradigm for light field angular super enhancement is presented without the need for ground truth. Following the learning paradigm, the learning models can be directly supervised by the input, and training data are not required to be paired as input and labels. Hence, more light field images with redundant parallax information can be used for the learning of deep light field reconstruction. To take advantage of the learning paradigm, A convolutional network leveraging motion estimation is built to synthesize novel views via fusing adjacent views. The experiments demonstrate that the method, implemented under the proposed learning paradigm, achieves high-quality metrics for simulated and real-world light field data. This is particularly notable for scenes that include multi-depth targets, complex textures, and large baselines. More accurate parallax structures can be recovered based on the proposed learning paradigm, whilst over 69% of training data are saved compared with other methods. In addition, under the proposed learning paradigm, even a simple shallow network can synthesize high-quality novel views. The PSNR achieved by the baseline method is approximately improved by 2dB after the proposed semi-supervision. Hence, the proposed semi-supervised learning paradigm can be easily integrated with other learning models. Regarding resolution enhancement of autostereoscopic measuring data, a self super-resolution algorithm driven by deep learning models has been designed. This algorithm is integrated into the measurement system, resulting in the creation of a self super-resolution autostereoscopic 3D measuring system. The self super-resolution algorithm can generate novel perspectives between the neighbouring Elemental Images (EIs) so that the angular resolution is markedly enhanced several times over. The proposed algorithm has been embedded into an autostereoscopic 3D measuring system so that the system can achieve self super-resolution. To validate the feasibility and technical merit of the proposed self super-resolution 3D measuring system, a comprehensive comparison experiment was conducted between the traditional autostereoscopic measuring system without super-resolution and the proposed system. The results demonstrate that the self super-resolution system can significantly improve the resolution of the measuring data by around four-fold and enhance the measurement accuracy with lower standard deviations and biases. To reduce the effect of vibration during on-machine measurement, multiple frames captured by the autostereoscopic measuring system are able to be used to eliminate the measurement errors induced by the vibration. Furthermore, essential information for achieving high spatial resolution in the measurement data can be extracted from the redundant subpixel-level information. As a result, the study introduces a multi-frame autostereoscopic system designed specifically for the on-machine measurement of three-dimensional surfaces, aimed at enhancing resolution. It leverages the vibrations produced by the machine tool during on-machine measurements to capture multiple frames of the target surface with offsets. This approach allows for resolution enhancement. A multi-frame resolution-enhanced deep learning model is developed, along with a supervised training process, to generate resolution-enhanced raw elemental images. This approach is pivotal to improving the measurement resolution. Through experiments, the system performance is assessed, and the results demonstrate a four-fold enhancement in spatial resolution along with improved measurement accuracy. In this study, learning-based techniques are applied to enhance LF resolution of measurement data gathered using the autostereoscopic 3D measuring system. Through experimental evaluation, the measurement resolution and accuracy for micro-structured surfaces are improved after artificial intelligence enhancement. In addition to the aforementioned advancements, the study also presents a generic semi-supervised learning paradigm specifically designed for deep learning models employed in angular resolution-enhancement tasks. This innovative paradigm allows for high data efficiency, ensuring superior performance in enhancing angular resolution using limited labelled data. The deep learning-based method results in an enhancement of angular resolution from 16 × 9 to 31 × 17, as well as an improvement in spatial resolution from 151 × 151 to 604 × 604. This enhancement in angular resolution leads to a reduction in error between measured and true values from over 1 µm to around 0.1 µm on average, along with a decrease in repeated measurement deviation by around 1 µm. Additionally, the spatial enhancement contributes to an increase in accuracy by 1 µm and a reduction in the deviation of repeated measurements from 1.533 µm to 1.388 µm. The research highlights the potential of combining autostereoscopy technology with deep learning technology for precise measurement. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13504