| 摘要: | 近年來人臉偵測 (face detection) 與人臉辨識 (face recognition) 技術被廣泛地應用在各種實務或商業娛樂系統上，像是門禁系統、監控系統、身分認證的登入系統、社群網站等。然而，人臉偵測率與人臉辨識率常常受到許多因素影響，包含光照環境的不同、表情的不同、臉部旋轉、及有遮蔽物的情形等。
本論文透過找出人臉的多尺度區域二元模式特徵 (Multi-scale Block Local Binary Pattern, MB-LBP) 來做人臉偵測，克服光線變化、模糊與雜訊，加入多角度人臉影像來訓練分類器，使得可以克服某些程度之內的臉部旋轉。並結合深度卷積神經網路 (deep convolutional neural network) 技術，用我們自己蒐集的多種光源、角度、清晰度變化的人臉資料庫，透過多層網路神經元架構來訓練學習人臉辨識，且與傳統的辨識方法做比較。
在實驗分析中，我們以自己拍攝的影片做測試 (包含不同光線變化，不同角度的人臉影像)。偵測方面，依照不同參數的調整，偵測率可以達到 91% ~ 97%, 誤判率 4 × 10^(-7), 並比較不同參數搭配的結果。辨識方面，與傳統方法相比，利用深度卷積神經網路方法訓練 12 個類別 (分別為 1 ~ 10 類、其他人和非人臉)的人臉辨識模型，測試樣本 881 個，辨識正確率可達到 94.3%。;In recent years, face recognition and face detection techniques are widely used in various applications, such as access control systems, surveillance system, login system, and community websites etc. However, there are some factors that affect the recognition performance like different lighting conditions, facial expression, face rotation, and occlusion by other objects.
In this paper, we use Multi-scale Block Local Binary Pattern (MB-LBP) to detect face. MB-LBP can overcome different lighting conditions, blurred and noise images. We add multi-angle face images to train the detection classifier, so that classifier can overcome some degree of face rotations. We collect face databases with various lighting conditions, angles, multi-resolution and use multi-layer neural network to train the face recognition system. We compare traditional recognition method with deep convolutional neural network (CNN).
In the experimental analysis, we do test with the videos which we shot, including different lighting conditions and different face angles. In the detection section, according to different parameters, the detection rate can reach up to 91% ~ 97% and about 4 × 10^(-7) false positive rate. We compare the results with different parameters. In the recognition section, comparing with traditional methods, we use 12 classes, including 10 persons, other men, and not men, to train deep convolutional neural network model for face recognition. In the case of 881 test samples, the recognition rate reach to 94.3%. |
