三維人臉模型合成:人臉辨識補助

Abstract

人臉辨識一直是個熱門的研究領域,並且有廣泛的應用場合;然 而,在現實的應用下仍面臨了許多挑戰。其中一點是彩色照片辨識系 統準確度在不同光線環境下的表現並不穩定,也就是當實際應用的場 合環境光線與測試資料的環境光線不同時,辨識系統的準確度可能會 比預期的還要差許多。相較於彩色照片,深度圖和紅外影像在不同光 線環境下所拍攝出來的結果則較為穩定,因此,本論文著重在利用深 度圖和紅外影像的人臉辨識。 本論文的目標為利用合成出的三維人臉模型增強人臉辨識系統的準 確度和尋找適當的模型參數。論文首先介紹如何使用多視角深度圖和 彩色圖合成出完整的人臉模型,接下來提出三個合成人臉模型有所幫 助的使用情境。第一個例子是利用合成人臉模型做為訓練實例,並以 提升模型準確度。第二個例子是合成未知人臉模型,用以輔助挑選開 放式人臉識別模型的參數。第三個例子是利用合成人臉模型投影出多 視角的紅外影像,做為訓練資料,以提升模型準確度。除了以上使用 情境外,最後也提出了一些合成人臉模型可能有所幫助的應用方向。

Face recognition has been a hot research area for its wide range of applications; however, there are still many challenges in practice. One of the challenges is that the color image recognition system is not stable under different illumination conditions, that is, the performance of a face recognition system might be worse than the expectation in different environments. There- fore, this thesis focuses on depth maps and infrared images recognition system because depth maps and infrared images are relatively stable for different lighting. The objective of this thesis is to use synthetic 3D facial models to en- hance the performance of face recognition system and select a suitable model parameter. First, we introduce the processing of 3D facial model synthesis, then propose three use cases which show that synthetic 3D facial models are useful. The first case is instance-based depth map recognition by using synthetic 3D facial models for training instances. The second case is unknown 3D facial model synthesis for the aids of the open-set face identification. We use synthetic unknown 3D facial models to predict the distance distribution of true unknown data. Lastly, the third case is multi-view infrared images synthesis using synthetic 3D facial model. In addition to these three cases, there are other applications that synthetic 3D facial model might be useful.