具有語義引導的特徵學習並應用於視覺化影像生成

Semantics-Guided Representation Learningwith Applications to Visual Synthesis

顏嘉緯(Jia-Wei Yan)

指導教授 ： 王鈺強

國立臺灣大學/電機資訊學院/電信工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201904354

摘要

學習可解釋的特徵表示一直是感興趣的主題之一。 大多數現有的研究不能容易地生成或操縱特徵表示，並通過插值生成具有特定語義的圖像。 在本文中，我們提出了一個角度三元組鄰居函數（ATNL），它能夠導出其分佈與語義資訊匹配的潛在特徵表示。 利用ATNL引導的潛在特徵空間，我們進一步利用球面語義內插來生成語義變化的圖像。 我們對MNIST和CMU Multi-PIE數據集的實驗證實了我們的ATNL和球形語義內插對最近的表示學習模型的有效性和強大性。

關鍵字

影像生成 ； 特徵學習 ； 深度學習 ； 機器學習 ； 電腦視覺

並列摘要

Learning interpretable representations has been among the topics of interest. Most existing works cannot easily generate or manipulate latent representations which semantically match the images of interest via interpolation. In this paper, we propose an Angular Triplet-Neighbor Loss (ATNL), which is able to derive latent representations whose distribution would match the semantic information. With the latent space guided by ATNL, we further utilize spherical semantic interpolation for generating semantic warping of images. Our experiments on both MNIST and CMU Multi-PIE datasets confirm the effectiveness and robustness of our ATNL and spherical semantic interpolation over recent representation learning models.

並列關鍵字

Image generation ； Representation learning ； Deep learning ； Machine learning ； Computer vision

參考文獻

[1] D. Berthelot, C. Raffel, A. Roy, and I. Goodfellow. Understanding and improving interpolation in autoencoders via an adversarial regularizer. In Proceedings of the International Conference on Learning Representations (ICLR), 2019.

[2] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In Advances in Neural Information Processing Systems (NIPS), 2016.

[3] J. Deng, J. Guo, N. Xue, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. arXiv preprint arXiv:1801.07698, 2019.

[4] J. Donahue, P. Krähenbühl, and T. Darrell. Adversarial feature learning. In Proceedings of the International Conference on Learning Representations (ICLR), 2017.

[5] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,