Face Synthesis Using Disentangled Representation-Learning GAN
碩士 === 國立臺灣科技大學 === 機械工程系 === 107 === We propose the Disentangled Representation Learning on a Triple Component Generative Adversarial Network (TC-GAN), for handling cross-pose and cross-age face recognition. The proposed TC-GAN has improved state-of-the-art DR-GAN with the Spectral Normalization c...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Others |
Language: | zh-TW |
Published: |
2019
|
Online Access: | http://ndltd.ncl.edu.tw/handle/d4j33g |
Summary: | 碩士 === 國立臺灣科技大學 === 機械工程系 === 107 === We propose the Disentangled Representation Learning on a Triple Component Generative Adversarial Network (TC-GAN), for handling cross-pose and cross-age face recognition. The proposed TC-GAN has improved state-of-the-art DR-GAN with the Spectral Normalization considered in the discriminator so that the generative and the adversarial framework can be better trained, and divide semi-supervised discriminator into two weights-independence Classifier and Discriminator for better classification capability. We also highlight the influences of pose/age imbalanced training data on the disentangled facial representation learning, and point out the difficulty of generating faces of extreme poses. We have explored the recently proposed nonlinear 3D Morphable Model (3DMM) and Conditional Adversarial Autoencoder (CAAE) to augment the training data, and verify the contributions made by the learning about augmented data. Experiments on the Multi-PIE and MORPH demonstrate the superiority of TC-GAN over other state-of-the-art approaches.
|
---|