階層式人臉偵測與對齊的平行化訓練

本論文目標為重現微軟亞洲研究院的陳先生等人在 2014 年提出的階層式人臉偵測與對齊 (JDA) 方法。我們為了提升 JDA 的效能而做了三種修改：首先，在特徵提取的過程中使用局部座標系，來改善影像特徵的選轉不變性；第二，更改產生非人臉訓練資料的方式，由更多影像中取得非人臉影像來增加訓練資料的多樣性；第三，使用 OpenMP 和多台機器，平行化 JDA 的訓練過程。最後透過 FDDB, CVF, CelebA 三個臉部影像資料集來檢測開發出的 JDA 模型。實驗結果顯示我們所發展出來的 JDA 模型性能優於網路上其他人重現的 JDA，且平行化訓練過程可以有效縮減訓練時間。此外，本論文也分析了為何重現的JDA模型無法達到原論文上顯示的精確度。

關鍵字

人臉偵測；人臉對齊；結合人臉特徵定位與偵測；分類迴歸樹；平行化運算；隨機森林迴歸分析與分類

並列摘要

This thesis aims to implement the technique of joint cascade face detection and alignment (JDA) proposed by Chen {it et al.} at Microsoft Research Asia in 2014. Three modifications are made to improve performance of JDA. First, a local coordinate system is introduced into the feature extraction process to improve the rotation invariance of the image feature. Second, the negative sample extraction process is modified to increase the diversity of negative samples by including more non-face images. Third, the JDA training process is parallelized by using OpenMP and multiple computers. The developed JDA model is tested by using three facial image data sets, namely, FDDB, CVF, and CelebA. Experimental results show that our JDA model outperforms the other JDA implementations available on the Internet. Furthermore, the parallelized training process can reduce the training time considerably. Additionally, it is analyzed why the implemented JDA model is not as accurate as the model of the original paper.

並列關鍵字

Face Detection ； Face Alignment ； Joint Detection and Alignment ； Classification and Regression Tree ； Parallel Computing ； Random Forest Regression/Classification