一种改进的Pre-Net网络及其在材质建模中的应用
An Improved Pre-Net Network and Its Application in Material Modeling
摘要: 材质建模是图形学研究与应用中的热点问题，目的是通过建模确定材质表观的纹理、法向、反射率和粗糙度等属性，从而提高对物体渲染的真实感。为了解决传统的材质建模需要耗费大量人力物力和时间，目前有人提出利用深度学习的方式实现材质的建模。本文在Xiao Li 提出的Pre-Net的基础上，对过拟合及预测精度问题进行研究，提出了改进模型的方法，通过添加Dropout层和Inception结构，从而减少过拟合现象并提高了模型的预测精度，同时使用批渲染技术快速大量地获取到实验所需的训练数据。实验表明，本文方法可以使模型的预测结果更加准确，泛化能力也更强。
Abstract: Material modeling is a hot issue in the research and application of graphics. The purpose is to de-termine the properties of material appearance such as texture, normal, reflectivity and roughness through modeling, so as to improve the realism of object rendering. In order to solve the problem of traditional material modeling, it takes a lot of manpower, resources and time. Based on the prenet proposed by Xiao Li , this paper studied the over-fitting and prediction accuracy, and proposed a method to improve the model. By adding Dropout layer and Inception structure, the over-fitting phenomenon was reduced and the prediction accuracy of the model was improved. At the same time, a large number of training data needed for the experiment were obtained quickly by using batch rendering technology. Experiments show that this method can make the prediction result of the model more accurate and the generalization ability stronger.
文章引用：杜金莲, 孔祥栋. 一种改进的Pre-Net网络及其在材质建模中的应用[J]. 计算机科学与应用, 2020, 10(4): 795-810. https://doi.org/10.12677/CSA.2020.104083

参考文献

[1] Xiao, L., Yue, D., Pieter, P. and Xin, T. (2017) Modeling Surface Appearance from a Single Photograph Using Self-Augmented Convolutional Neural Networks. ACM Transactions on Graphics, 36, 1-45, 11. https://dl.acm.org/doi/abs/10.1145/3072959.3073641 [Google Scholar] [CrossRef] 
[2] Sean, B., Paul, U., Noah, S. and Kavita, B. (2013) OpenSurfaces: A Richly Annotated Catalog of Surface Appearance. ACM Transactions on Graphics, 32, 1-111, 17. [Google Scholar] [CrossRef] 
[3] Dror, R.O., Adelson, E.H. and Willsky, A.S. (2001) Recognition of Surface Reflectance Properties from a Single Image under Unknown Real-World Illumination. Proceedings of IEEE Workshop on Identifying Objects across Variations in Lighting: Psychophysics and Computation. http://dspace.mit.edu/bitstream/handle/1721.1/6664/AIM-2001-033.pdf?sequence=2
[4] Zhang, R., Zhu, J.-Y., Isola, P., Geng, X.Y., Lin, A.S., Yu, T.H. and Efros, A.A. (2017) Real-Time User-Guided Image Colorization with Learned Deep Priors. ACM Transactions on Graphics, 36, 4. [Google Scholar] [CrossRef] 
[5] Ren, P.R., Wang, J.P., Snyder, J., Tong, X. and Guo, B.N. (2011) Pocket Reflectometry. ACM Transactions on Graphics, 30, 4. https://dl.acm.org/doi/abs/10.1145/1964921.1964940 [Google Scholar] [CrossRef] 
[6] Oxholm, G. and Nishino, K. (2012) Shape and Reflectance from Natural Illumination. ECCV, 528-541. [Google Scholar] [CrossRef] 
[7] Aittala, M., Aila, T. and Lehtinen, J. (2016) Reflectance Modeling by Neural Texture Synthesis. ACM Transactions on Graphics, 35, 1-65, 13. https://dl.acm.org/doi/abs/10.1145/2897824.2925917 [Google Scholar] [CrossRef] 
[8] Matusik, W., Pfister, H., Brand, M. and McMillan, L. (2003) Effi-cient Isotropic BRDF Measurement. In Proceedings of the 14th Eurographics Workshop on Rendering (EGRW’3). Eurographics Association, Aire-la-Ville, Switzerland, Switzerland, 241-247.
https://dl.acm.org/doi/abs/10.5555/882404.882439
[9] Wang, J.P., Zhao, S., Tong, X., Snyder, J. and Guo, B.N. (2008) Modeling Anisotropic Surface Reflectance with Example-Based Microfacet Synthesis. ACM SIGGRAPH 2008 Papers (SIGGRAPH’08). ACM, New York, NY, Article 41. https://dl.acm.org/doi/abs/10.1145/1399504.1360640 [Google Scholar] [CrossRef] 
[10] Zhou, Z.M., Chen, G.J., Dong, Y., Wipf, D., Yu, Y., Snyder, J. and Tong, X. (2016) Sparse-as-Possible SVBRDF Acquisition. ACM Transactions on Graphics, 35, Article 189. https://dl.acm.org/doi/abs/10.1145/2980179.2980247 [Google Scholar] [CrossRef]