本論文提出一個基於卷積類神經網路(Convolutional Neural Network)的方法進行三維臉部模型生成以用於資料擴增並且實現三維臉部識別的方法。過去幾年來類神經網路在二維臉部辨識上取得重大成就,例如VGG (Visual Geometry Group) Face、Inception和ResNet (Residual Network)。這些網路有含有大量參數必須由非線性最佳化的方法來調整,因此就需要大量的訓練資料來調整。2017年開始,蘋果電腦推出iPhone X智慧型手機,其中FaceID技術把人臉識別技術由二維推向三維,三維人臉辨識成為風潮。然而要訓練三維人臉辨識並不容易,首先訓練資料非常稀少,最大的三維人臉資料集中,也只有數千張人臉的深度圖,並且只有數百個個體。對此,本論文中使用遷移學習(Transfer Learning)技術來應對這個困難,並且藉由生成三維臉部模型增加訓練資料的歧異度與數量以增強三維臉部識別效能。
A method of data augmentation for 3D face model and using it for 3D face identification is proposed in this thesis. In the past few years, researchers have achieved significant progress on 2D face identification and verification through neural network approaches, such as VGG (Visual Geometry Group) Face, GoogleNet Inception, and ResNet (Residual Network). Since there are so many hyper parameters that need to be optimized in neural networks, large data must be provided for training. In 2017, FaceID was proposed by Apple Inc. Face identification has been scaled up from 2D to 3D. However, training a 3D face classifier is difficult. 3D face datasets nowadays are so small that even a large set of 3D face (Bosphorus 3D Face Dataset) contains only 4,666 faces of 105 identities. In order to solve the lack of data, we use transfer learning [13], and several data augmentation methods by generating face mesh from different views to make the classifier more robust and discriminative.