Векторизация изображений с помощью глубокого обучения тема диссертации и автореферата по ВАК РФ 00.00.00, кандидат наук Егиазарян Ваге Грайрович

• Егиазарян Ваге Грайрович
• кандидат наук
• 2024

Оглавление диссертации кандидат наук Егиазарян Ваге Грайрович

Contents

1 Dissertation Topic

2 Key Results and Contributions

3 Publications and Validation of the Work

4 Dissertation Structure

5 Description of Data and Problem Statement

5.1 Description of Data for Two-Dimensional Drawings

5.2 Description of Data for Three-Dimensional Objects

6 Vectorization of Two-Dimensional Images

6.1 Preprocessing of images

6.2 Initial primitives estimation

6.3 Primitive parameters optimization

6.4 Experimental evaluation

7 Generation of High-Quality Synthetic Data with Descriptions of Three-Dimensional Objects

8 Vectorization of Three-Dimensional Objects

8.1 Description of the Developed Method for Obtaining Three-Dimensional Parametric Curves

8.1.1 Initialization

8.1.2 Corner Detection

8.1.3 Segmentation of curves and angles

8.1.4 Curve Graph Extraction

8.1.5 Spline Approximation and Optimization

8.1.6 Post-processing of Spline-Based Representations

8.2 Evaluation of the Proposed Approach

9 Conclusion 34 References 35 Appendix A Article 1: Deep Vectorization of Technical Drawings

Appendix B Article 2: DEF: Deep Estimation of Sharp Geometric Features in 3D Shapes

Appendix C Article 3: Latent-Space Laplacian Pyramids for Adversarial Representation Learning with 3D Point Clouds

Appendix D Russian Translation of the Ph.D. dissertation

Введение диссертации (часть автореферата) на тему «Векторизация изображений с помощью глубокого обучения»

1 Dissertation Topic

The theme of this research is the development of efficient methods for raster image and depth map for three-dimensional object vectorization using deep learning. Vectorization of objects refers to finding object representation using mathematical primitives and relationships between them.

To achieve this goal, the following tasks were addressed: data collection, construction of mathematical models, and development of vectorization algorithms. Data collection was performed through synthetic data generation and processing of real scanned images of two-dimensional and three-dimensional objects. Automatic algorithms using computer vision methods were developed for data cleaning and processing, along with procedures for manual data processing. The proposed algorithms enable obtaining annotated data in a semi-automatic mode, which opens up the possibility of using deep learning methods to train neural networks on those data.

Various neural network architectures, such as convolutional and recurrent networks, as well as transformers, were investigated to build architectures that effectively solve the set tasks. The aim of constructing an optimal architecture is to create models capable of accurately and efficiently vectorizing various types of images and three-dimensional objects.

The proposed algorithms demonstrate high accuracy and efficiency in solving object vec-torization tasks. These algorithms can be applied in various fields such as computer vision, robotics, and data visualization to represent objects as vector graphics with high accuracy and detail.

The relevance of research

There are many software products for image and document recognition, but most of them focus on text recognition and either do not work with images or perform poorly with them. Typically, these products do not allow representing the recognition results as vector graphics, one of the most convenient formats for subsequent work. In the case of two-dimensional and three-dimensional data, existing vectorization approaches based solely on optimization algorithms have a large number of customizable hyperparameters and perform poorly with noise in the input data. Using neural networks can reduce the number of hyperparameters, and because neural networks learn to recognize patterns, it is possible to efficiently process even significantly noisy data.

The relevance of the study is justified by the fact that image and form vectorization is an important and complex task in the field of computer vision, which has not been fully solved yet. In this dissertation, the problem of precise vectorization for two-dimensional technical drawings and three-dimensional component models is considered, and a method for recognizing drawings and objects using neural networks is proposed for tasks requiring precise and flexible representation of two-dimensional and three-dimensional objects. The research results can be

applied in various applications such as automatic recognition and analysis of technical drawings, automated modeling and editing of 3D objects, as well as virtual and augmented reality.

The objective of this dissertation is to study and develop efficient methods for vectoriza-tion of two-dimensional technical drawings and three-dimensional models in the field of computer vision. The main tasks include studying existing methods, collecting and processing data, developing deep learning algorithms, evaluating and comparing vectorization methods, and studying their practical application.

The objective is achieved through the solution of the following tasks:

• obtaining high-quality geometric data for two-dimensional and three-dimensional objects for subsequent training of neural network models;

• developing a new system for vectorization of raster images of technical drawings;

• developing an algorithm for reconstructing parametric models of three-dimensional objects.

The work is based on the use of methodology and methods of machine learning, differential geometry, and mathematical modeling.

2 Key Results and Contributions

The work is based on the use of methodology and methods of computer vision, computer graphics, machine learning, and deep learning.

The scientific novelty of the work can be described by the following points:

1. A new method was proposed for generating high-resolution three-dimensional objects at different scales. A new deep multi-scale model for point cloud generation, Latent-Space Laplacian Pyramid GAN, was developed based on advanced technologies in the field of generative adversarial networks for point cloud data [1] and approaches for modeling data at different scales [8, 26].

2. For the first time, a system was proposed that allows for a more precise solution to the vectorization problem of scanned drawings. The system includes several components: specially prepared high-quality data for training neural network models, a new architecture of the trainable neural network model for vectorization and a new approach for optimizing primitives to build the final representation of the object in vector format. Train-able neural network models are used at multiple steps of this approach: for preprocessing the image to reduce noise levels; for obtaining an initial vector approximation of the image. Then, using the solution to an optimization task, the final result is constructed from the obtained initial approximation, and its post-processing is performed to minimize the number of primitives.

3. A new method for extracting parametric curves from three-dimensional point clouds was presented, allowing for accurate transformation of point clouds into analytical models of special curves in three-dimensional space, which describe the structural features of three-dimensional component models and are necessary for further building 3D descriptions of these objects.

These results were published in the proceedings of international conferences at the Core A* and Core B level, with the papers undergoing double-blind peer review, and the presentations being given at international conferences.

Theoretical and Practical Significance

The theoretical significance of the work lies in the new methods and algorithms in the field of image processing, vectorization, and generation of three-dimensional data. The current state of approaches to vectorization and generation of images and three-dimensional data was analyzed in the study. Advanced approaches from machine learning, modern neural network architectures and optimization procedures were applied to build models and methods. Following the obtained practical results the understanding of the applicability boundaries of modern machine learning methods for vectorization tasks has been expanded.

The developed methods have practical applicability in various fields, such as automatic recognition and analysis of technical drawings, automated modeling and editing of three-dimensional objects.

Results Presented for Defense:

1. Methodology and algorithms for data generation,

2. Algorithm for transforming raster technical drawings into vector graphics while preserving information,

3. Algorithm for transforming three-dimensional scanned objects into vector graphics consisting of three-dimensional curves,

4. Methodology for evaluating the accuracy and quality of vectorized data.

The reliability of the obtained results is ensured by the correct use of a tested set of tools for research and analysis. The proposed algorithms were experimentally tested on various tasks and on real datasets of both two-dimensional and three-dimensional objects. Detailed descriptions of the experiments conducted and the source code are publicly available, enabling replication of the obtained results. The research results were published in leading scientific journals and at conferences dedicated to computer vision and pattern recognition.

Personal Contribution to the Presented Results: All stated results were obtained by the author of this dissertation. In all mentioned cases, both the text of the papers and the experimental results presented therein are the results of collaboration with other authors.

In Paper 1: "Deep vectorization of technical drawings," Egiazaryan V.G., as the primary author, developed and implemented a system for processing input images with subsequent acquisition of vector representations. Additionally, the author was responsible for developing the neural network architecture and its training algorithm, which transforms raster images into vector primitives, and jointly with co-authors developed methods and experimental design for evaluating the effectiveness of the proposed method.

In Paper 2: "Def: Deep estimation of sharp geometric features in 3d shapes," Egiazaryan V.G. was responsible for developing and improving the second component of the entire pipeline: the algorithm for recovering parametric curves from point clouds.

In Paper 3: "Latent-Space Laplacian Pyramid GAN," the author of the dissertation developed and implemented the algorithm for multi-scale generation of three-dimensional point clouds, as well as a methodology for evaluating the accuracy and quality of the generated objects.

3 Publications and Validation of the Work

First-tier publications

1. Vage Egiazarian*, Oleg Voynov*, Alexey Artemov, Denis Volkhonskiy, Aleksandr Safin, Maria Taktasheva, Denis Zorin, Evgeny Burnaev. Deep vectorization of technical drawings. In Proceedings of the European Conference on Computer Vision, pp. 582-598, Springer, Cham. ECCV 2020, CORE A*, indexed by SCOPUS.

2. Albert Matveev*, Ruslan Rakhimov*, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, Evgeny Burnaev. DEF: Deep estimation of sharp geometric features in 3d shapes. In Proceedings of the SIGGRAPH 2022 conf., ACM Transactions on Graphics (TOG), 41(4):1-22, 2022. CORE A*, indexed by SCOPUS.

Second-tier publications

3. Vage Egiazarian*, Savva Ignatyev*, Alexey Artemov, Oleg Voynov, Andrey Kravchenko, Youyi Zheng, Luiz Velho, Evgeny Burnaev. Latent-space laplacian pyramids for adversarial representation learning with 3d point clouds. In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pp. 421-428, VISIGRAPP 2020, CORE B, indexed by SCOPUS.

The author has also contributed to the following publication

4. Oleg Voynov, Alexey Artemov, Vage Egiazarian, Alexander Notchenko, Gleb Bobrovskikh, Denis Zorin, Evgeny Burnaev. Perceptual deep depth super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 5653-5663,2019. CORE A*, indexed by SCOPUS.

5. Denis Mazur*, Vage Egiazarian*, Stanislav Morozov*, Artem Babenko. Beyond vector spaces: Compact data representation as differentiable weighted graphs. In Proceedings of the Advances in Neural Information Processing Systems (NeurIPS), Volume. 32, 2019. CORE A*, indexed by SCOPUS.

6. Alexander Korotin, Vage Egiazarian, Arip Asadulaev, Alexander Safin, Evgeny Burnaev. Wasserstein-2 generative networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. CORE A*, indexed by SCOPUS.

7. Alexander Korotin, Vage Egiazarian, Lingxiao Li, Evgeny Burnaev. Wasserstein iterative networks for barycenter estimation. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2022. Volume 35. CORE A*, indexed by SCOPUS.

8. Arip Asadulaev*, Alexander Korotin*, Vage Egiazarian, Petr Mokrov, Evgeny Burnaev. Neural optimal transport with general cost functionals. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. CORE A*, indexed by SCOPUS.

9. Tim Dettmers*, Ruslan Svirschevski*, Vage Egiazarian*, Denis Kuznedelev*, Elias Frantar, Saleh Ashkboos, Alexander Borzunov, Torsten Hoefler, Dan Alistarh. SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression. In Proceedings of the International Conference on Learning Representations (ICLR), 2024. CORE A*, indexed by SCOPUS.

10. Vage Egiazarian*, Andrei Panferov*, Denis Kuznedelev, Elias Frantar, Artem Babenko, Dan Alistarh. Extreme Compression of Large Language Models via Additive Quantization. In Proceedings of The Forty-first International Conference on Machine Learnings (ICML), 2024. CORE A*, indexed by SCOPUS.

Reports at conferences and seminars

• "Deep vectorization of technical drawings", poster presentation at 16th European Conference on Computer Vision (ECCV), Online, 2020;

• "Vectorization using deep learning," presentation at the seminar of the Association "Artificial Intelligence in Industry," Online;

• "DEF: Deep estimation of sharp geometric features in 3D shapes," poster presentation at the SIGGRAPH conference, Canada, 2022;

• "Latent-space Laplacian pyramids for adversarial representation learning with 3D point clouds," poster presentation at the VISAPP conference, Malta, 2020.

* - Equal contribution

4 Dissertation Structure

The main part of the dissertation consists of four sections:

Section 5 describes the methodology for obtaining vectorization data. Methods for obtaining both synthetic and real data necessary for training and evaluating vectorization models are considered.

Section 6 presents a data processing system for 2D technical drawings aimed at obtaining vector graphics. Methods and algorithms used to transform raster images into vector representations are described.

Section 7 describes a new method for generating high-quality three-dimensional objects. Existing methods for generating three-dimensional data are discussed, and the general architecture and structure of the proposed new method are described.

Section 8 proposes a method for obtaining three-dimensional parametric curves from three-dimensional objects. The efficiency of the developed method is compared to existing approaches.

Заключение диссертации по теме «Другие cпециальности», Егиазарян Ваге Грайрович

Заключение

В данном разделе был представлен метод векторизации изображений глубины для получения параметрического представления трехмерного объекта. Здесь векторным представлением трехмерного объекта служит каркас объекта, представленный с помощью параметрических кривых.

Для того чтобы получить параметрические кривые, на каждое из входных изображений глубины была применена нейронная сеть на основе U-Net для предсказания расстояния до особых линий. Далее полученный результат был объединен в трехмерное поле расстояний. На основе этого поля были извлечены параметрические кривые, и таким образом было получено векторное представление входного объекта.

9. Заключение

Данное исследование посвящено разработке методов для решения задач векторизации RGB изображений и изображений глубины с использованием глубокого обучения. Для решения данной задачи были разработаны и применены алгоритмы глубокого обучения и оптимизации, специально адаптированные для задач векторизации двухмерных технических чертежей и трехмерных моделей. Были собраны и предобработаны соответствующие данные, включая синтетически сгенерированные и реальные изображения и модели. Это позволило создать наборы данных, необходимые для тренировки и оценки разработанных моделей. Затем была произведена разработка нейронных сетей и методов оптимизации, которые позволяют точно и эффективно векторизировать входные объекты.

В рамках данного исследования были получены следующие результаты:

• предложен метод получения высококачественных геометрических данных для двухмерных и трехмерных объектов;

• разработана новая система для векторизации растровых изображений технических чертежей;

• разработан алгоритм для восстановления параметрических моделей, описывающих особые кривые у трехмерных форм.

Результаты данного исследования представляют собой значимый вклад в область глубокого обучения и векторизации объектов. С помощью разработанных алгоритмов была достигнута высокая точность и эффективность при решении задач векторизации объектов. Представленные модели демонстрируют способность извлекать математические примитивы и связи между ними, представляя объекты в виде точных векторных представлений. Разработанные алгоритмы имеют большой потенциал для применения в различных областях, включая распознавание и анализ технических чертежей, автоматизированное моделирование и редактирование трехмерных объектов.

В дальнейших исследованиях можно углубиться в изучение альтернативных архитектур нейронных сетей, разработку более сложных оптимизационных методов и рассмотрение других типов данных для векторизации. Также важно продолжать работу над улучшением качества и обобщающей способности моделей.

Данное исследование является важным шагом в направлении разработки эффективных методов векторизации объектов с использованием глубокого обучения. Разработанные методы открывают новые перспективы для различных приложений и дальнейших исследований, стимулируя развитие области в целом и внося ощутимый вклад в научное сообщество.

Список литературы диссертационного исследования кандидат наук Егиазарян Ваге Грайрович, 2024 год

Список литературы

[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, pages 40--49,2018.

[2] Bin Bao and Hongbo Fu. Vectorizing line drawings with near-constant line width. In 2012 19th IEEE International Conference on Image Processing, pages 805--808. IEEE, 2012.

[3] Mikhail Bessmeltsev and Justin Solomon. Vectorization of line drawings via polyvector fields. ACM Transactions on Graphics (TOG), 38(1):9, 2019.

[4] Jiazhou Chen, Mengqi Du, Xujia Qin, and Yongwei Miao. An improved topology extraction approach for vectorization of sketchy line drawings. The Visual Computer, 34(12):1633--1644, 2018.

[5] JiaZhou Chen, Qi Lei, YongWei Miao, and QunSheng Peng. Vectorization of line drawing image based on junction analysis. Science China Information Sciences, 58(7):1--14, 2015.

[6] Xuelin Chen, Baoquan Chen, and Niloy J Mitra. Unpaired point cloud completion on real scans using adversarial training. arXiv preprint arXiv:1904.00069,2019.

[7] Robert L. Cook. Stochastic sampling in computer graphics. ACM Trans. Graph., 5(1):51-72, jan 1986.

[8] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In NIPS, pages 1486--1494,2015.

[9] Luca Donati, Simone Cesano, and Andrea Prati. A complete hand-drawn sketch vectorization framework. Multimedia Tools and Applications, 78(14):19083--19113,2019.

[10] Vage Egiazarian, Savva Ignatyev, Alexey Artemov, Oleg Voynov, Andrey Kravchenko, Youyi Zheng, Luiz Velho, and Evgeny Burnaev. Latent-space laplacian pyramids for adversarial representation learning with 3d point clouds. In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. SCITEPRESS - Science and Technology Publications, 2020.

[11] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Advances in neural information processing systems, pages 6059--6068, 2018.

[12] Jean-Dominique Favreau, Florent Lafarge, and Adrien Bousseau. Fidelity vs. simplicity: a global approach to line drawing vectorization. ACM Transactions on Graphics (TOG), 35(4):120, 2016.

[13] Jun Gao, Chengcheng Tang, Vignesh Ganapathi-Subramanian, Jiahui Huang, Hao Su, and Leonidas J Guibas. Deepspline: Data-driven reconstruction of parametric curves and surfaces. arXiv preprint arXiv:1901.03781, 2019.

[14] Yi Guo, Zhuming Zhang, Chu Han, Wen-Bo Hu, Chengze Li, and Tien-Tsin Wong. Deep line drawing vectorization via line subdivision and topology reconstruction. Comput. Graph. Forum, 38:81-90, 2019.

[15] David Ha and Douglas Eck. A neural representation of sketch drawings. In International Conference on Learning Representations, 2018.

[16] JH Hannay and JF Nye. Fibonacci numerical integration on a sphere. Journal of Physics A: Mathematical and General, 37(48):11591, 2004.

[17] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2 edition, 2004.

[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770--778, 2016.

[19] Xavier Hilaire and Karl Tombre. Robust and accurate vectorization of line drawings. IEEE Transactions on Pattern Analysis & Machine Intelligence, (6):890--904, 2006.

[20] Ruchin Kansal and Subodh Kumar. A vectorization framework for constant and linear gradient filled regions. The Visual Computer, 31(5):717--732,2015.

[21] Tapas Kanungo, Robert M. Haralick, Henry S. Baird, Werner Stuezle, and David Madi-gan. A statistical, nonparametric methodology for document degradation model validation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):1209--1223,2000.

[22] Byungsoo Kim, Oliver Wang, A Cengiz Oztireli, and Markus Gross. Semantic segmentation for line drawing vectorization using neural networks. In Computer Graphics Forum, volume 37, pages 329--338. Wiley Online Library, 2018.

[23] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9601--9611,2019.

[24] Chengze Li, Xueting Liu, and Tien-Tsin Wong. Deep extraction of manga structural lines. ACM Transactions on Graphics (TOG), 36(4):117,2017.

[25] Chun-Liang Li, Manzil Zaheer, Yang Zhang, Barnabas Poczos, and Ruslan Salakhutdinov. Point cloud gan. arXiv preprint arXiv:1810.05795,2018.

[26] Priyanka Mandikal and Venkatesh Babu Radhakrishnan. Dense 3d point cloud reconstruction using a deep pyramid network. In WACV, pages 1052--1060. IEEE, 2019.

[27] Albert Matveev, Alexey Artemov, Denis Zorin, and Evgeny Burnaev. 3d parametric wireframe extraction based on distance fields. In 2021 4th International Conference on Artificial Intelligence and Pattern Recognition, AIPR 2021, page 316-322, New York, NY, USA, 2021. Association for Computing Machinery.

[28] Albert Matveev, Ruslan Rakhimov, Alexey Artemov, Gleb Bobrovskikh, Vage Egiazarian, Emil Bogomolov, Daniele Panozzo, Denis Zorin, and Evgeny Burnaev. Def: Deep estimation of sharp geometric features in 3d shapes, 2022.

[29] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784,2014.

[30] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH 2021), 40(4):51:1--51:14, 2021.

[31] Vishaal Munusamy Kabilan, Brandon Morris, and Anh Nguyen. Vectordefense: Vectorization as a defense to adversarial examples. arXiv preprint arXiv:1804.08529, 2018.

[32] Patryk Najgebauer and Rafal Scherer. Inertia-based fast vectorization of line drawings. Comput. Graph. Forum, 38:203--213, 2019.

[33] Gioacchino Noris, Alexander Hornung, Robert W Sumner, Maryann Simmons, and Markus Gross. Topology-driven vectorization of clean line drawings. ACM Transactions on Graphics (TOG), 32(1):4, 2013.

[34] Open CASCADE Technology OCCT. https://www.opencascade.com/, 2021. Accessed: 2021-06-01.

[35] PrecisionFloorplan. PrecisionFloorplan. http://precisionfloorplan.com. Accessed: 2020-03-05.

[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234--241. Springer, 2015.

[37] Peter Selinger. Potrace: a polygon-based tracing algorithm. Potrace (online), http://po-trace.sourceforge.net/potrace.pdf (2009-07-01), 2003.

[38] Edgar Simo-Serra, Satoshi Iizuka, and Hiroshi Ishikawa. Mastering sketching: adversarial augmentation for structured prediction. ACM Transactions on Graphics (TOG), 37(1):11, 2018.

[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998--6008,2017.

[40] Oleg Voynov, Gleb Bobrovskikh, Pavel Karpyshev, Saveliy Galochkin, Andrei-Timotei Ardelean, Arseniy Bozhenko, Ekaterina Karmanova, Pavel Kopanev, Yaroslav Labutin-Rymsho, Ruslan Rakhimov, et al. Multi-sensor large-scale dataset for multi-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21392--21403,2023.

[41] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliasacchi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in Neural Information Processing Systems, 33, 2020.

[42] Jiaojiao Zhao, Jie Feng, and Bingfeng Zhou. Image vectorization using blue-noise sampling. In Imaging and Printing in a Web 2.0 World IV, volume 8664, page 86640H. International Society for Optics and Photonics, 2013.