Michigan State University
Representation learning and image synthesis for deep face recognition
Abstract
dc:description"Face recognition has been advanced a lot in recent years thanks to the development of deep neural networks. The large intra-class variations in pose, illumination, and expression are the long-standing challenges. Learning a discriminative representation that is robust to these variations is the key. In the scenarios of profile pose or long-tail training data, image or feature-level data augmentation is needed. This dissertation presents three different methods to solve these problems. First, we explore a multi-task Convolutional Neural Network (CNN) that aims to leverage side tasks to improve representation learning. A pose-directed multi-task CNN is introduced to better handle pose variation. The proposed framework is effective in pose-invariant face recognition. Second, we propose a Face Frontalization-Generative Adversarial Network (FF-GAN) that can generate a frontal face even from an input image with extreme profile pose. FF-GAN handles pose variation from the perspective of image-level data augmentation. Multiple loss functions are proposed to achieve large-pose face frontalization. The proposed approach is evaluated on various tasks including face reconstruction, landmark detection, face frontalization, and face recognition. Third, a feature transfer learning method is presented to solve the problem of insufficient intra-class variation via feature-level data augmentation. A Gaussian prior is assumed across all the regular classes and the variance are transferred from regular classes to long-tail classes. Further, an alternating training regimen is proposed to simultaneously achieve less biased decision boundaries and more discriminative representations. Extensive experiments have demonstrated the effectiveness of the proposed feature transfer framework."--Abstract.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Yin, Xi
- Contributors dc:contributor
-
- Liu, Xiaoming
- Jain, Anil K.
- Ross, Arun
- Morris, Daniel D.
- Michigan State University
Subjects
dc:subject × 2Rights
dc:rights- Statement dc:rights
-
- In Copyright
- Language dc:language
- English
Identifiers
dc:identifier.*- Identifier
-
etd:19404
isbn:9780438289192
isbn:0438289196
oclc:1078987923
umi:10930064
local:Yin_grad.msu_0128D_16300