人臉識別是近年來受關注的熱門科技之一，特別是在深度學習與硬體設備的幫助下，實用價值更提升、辨識精準度越高。其中，訓練資料集的數量與深度學習的準確度有高度的相關性，目前大部分知名的人臉識別模型都使用到百萬張以上的人臉影像作為訓練資料，此外資料品質、資料集的分布偏差也會影響模型學習的成效。然而，相較於二維人臉識別，深度學習在三維人臉識別的發展較受限，很大的原因在三維臉部資料集的缺乏。在此篇論文，我們嘗試使用針對三維人臉的資料擴充方法來提升三維人臉識別的穩健性。利用合成的大量虛擬三維人臉資料，我們在人臉表情、臉部角度做變化增加資料的多樣性，並且在實驗探討：使用虛擬合成的資料是否可以增加三維人臉識別的強健性？我們證實使用虛擬合成的人臉資料可以有效地幫助三維人臉識別系統。;In recent years, deep learning has important increased the performance of 2D face recognition systems with the use of large-scale labeled image data. Deep neural networks can be closely approaching human-level depend heavily on the amount and quality of facial training data. However, contrast with 2D face recognition, training discriminative deep features for 3D face recognition is very difficult. Because of the unavailability of large training datasets, recognition accuracies have already saturated on existing 3D face datasets due to their small gallery sizes. Unlike 2D photograph, the collection of annotated high-quality large 3D facial scan datasets cannot be sourced from the web. In this paper, we show that using synthetically generated data as CNN training dataset can effectively work for 3D face recognition by fine-tuning the CNN with real-world data. We propose a 3D augmentation method for enlarging 3D facial data, we can generate 3D facial data with arbitrary amounts of facial identities, facial expression and pose variations by using 3D morphable face model. Finally, in our experiment we use two real-world 3D facial datasets to be compared. Our method outperforms the 3D face recognition system training only with real-world dataset. As well as, we find the significant accuracy improvement with the help from synthetic 3D facial data.
