Chen, Danwu (2023)
Deep learning for characterizing full-color 3D printers: accuracy, robustness, and data-efficiency.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00026378
Ph.D. Thesis, Primary publication, Publisher's Version
Text
DanwuChen_PhDThesis_28.11.2023.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (55MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Deep learning for characterizing full-color 3D printers: accuracy, robustness, and data-efficiency | ||||
Language: | English | ||||
Referees: | Kuijper, Prof. Dr. Arjan ; Fellner, Prof. Dr. Dieter ; Urban, Prof. Dr. Philipp | ||||
Date: | 1 December 2023 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xii, 123 Seite | ||||
Date of oral examination: | 21 November 2023 | ||||
DOI: | 10.26083/tuprints-00026378 | ||||
Abstract: | High-fidelity color and appearance reproduction via multi-material-jetting full-color 3D printing has seen increasing applications, including art and cultural artifacts preservation, product prototypes, game character figurines, stop-motion animated movie, and 3D-printed prostheses such as dental restorations or prosthetic eyes. To achieve high-quality appearance reproduction via full-color 3D printing, a prerequisite is an accurate optical printer model that is a predicting function from an arrangement or ratio of printing materials to the optical/visual properties (e.g. spectral reflectance, color, and translucency) of the resulting print. For appearance 3D printing, the model needs to be inverted to determine the printing material arrangement that reproduces distinct optical/visual properties such as color. Therefore, the accuracy of optical printer models plays a crucial role for the final print quality. The process of fitting an optical printer model's parameters for a printing system is called optical characterization, which requires test prints and optical measurements. The objective of developing a printer model is to maximize prediction performance such as accuracy, while minimizing optical characterization efforts including printing, post-processing, and measuring. In this thesis, I aim at leveraging deep learning to achieve holistically-performant optical printer models, in terms of three different performance aspects of optical printer models: 1) accuracy, 2) robustness, and 3) data efficiency. First, for model accuracy, we propose two deep learning-based printer models that both achieve high accuracies with only a moderate number of required training samples. Experiments show that both models outperform the traditional cellular Neugebauer model by large margins: up to 6 times higher accuracy, or, up to 10 times less data for a similar accuracy. The high accuracy could enhance or even enable color- and translucency-critical applications of 3D printing such as dental restorations or prosthetic eyes. Second, for model robustness, we propose a methodology to induce physically-plausible constraints and smoothness into deep learning-based optical printer models. Experiments show that the model not only almost always corrects implausible relationships between material arrangement and the resulting optical/visual properties, but also ensures significantly smoother predictions. The robustness and smoothness improvements are important to alleviate or avoid unacceptable banding artifacts on textures of the final printouts, particularly for applications where texture details must be preserved, such as for reproducing prosthetic eyes whose texture must match the companion (healthy) eye. Finally, for data efficiency, we propose a learning framework that significantly improves printer models' data efficiency by employing existing characterization data from other printers. We also propose a contrastive learning-based approach to learn dataset embeddings that are extra inputs required by the aforementioned learning framework. Experiments show that the learning framework can drastically reduce the number of required samples for achieving an application-specific prediction accuracy. For some printers, it requires only 10% of the samples to achieve a similar accuracy as the state-of-the-art model. The significant improvement in data efficiency makes it economically possible to frequently characterize 3D printers to achieve more consistent output across different printers over time, which is crucial for color- and translucency-critical individualized mass production. With these proposed deep learning-based methodologies significantly improving the three performance aspects (i.e. accuracy, robustness, and data efficiency), a holistically-performant optical printer model can be achieved, which is particularly important for color- and translucency-critical applications such as dental restorations or prosthetic eyes. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-263785 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science | ||||
Divisions: | 20 Department of Computer Science > Interactive Graphics Systems | ||||
Date Deposited: | 01 Dec 2023 13:09 | ||||
Last Modified: | 05 Dec 2023 07:37 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/26378 | ||||
PPN: | 513631097 | ||||
Export: |
View Item |