9 A New Feature Set for Face Detection台湾清华大学的一篇论文.docx

1、9 A New Feature Set for Face Detection台湾清华大学的一篇论文國 立 清 華 大 學 碩 士 論 文 題目: 使用新特徵的人臉偵測系統A New Feature Set for Face Detection系別 資訊系統與應用研究所 組別 甲 學號姓名 926701蔡忠志 (Chung-Chih Tsai) 指導教授 張智星博士 (Jyh-Shing Roger Jang) 中華民國九十四年六月Abstract Viola and Jones introduce a fast face detection system which uses a cascad

2、ed structure that can achieve high detection rate and low false positive rate. Their system uses integral images to compute values from features. This thesis introduces two new types of integral images which are called triangle integral images and two corresponding features which are named triangle

3、features. And this thesis proposes a method to lower training error by modifying Discrete AdaBoost. As results, to use triangle features can decrease the numbers of features; this research achieves lower false positive rate and fewer features are used.摘要 在快速人臉偵測的研究中，Viola和Jones提出了一個連接式的架構，此架構能得到高辨識率

4、及低錯誤率；他們使用integral image來計算人臉特徵值。本研究提出了兩種新的integral image: triangle integral image以及各別對應的三角特徵。另外，本研究以Discrete AdaBoost為基礎，提出了一個能在訓練時降低非人臉錯誤率的方法。我們的實驗證明，三角特徵能使得需要的feature數減少；改進過後的AdaBoost能使得錯誤率更低。Keywords Face detection, integral image, feature, AdaBoost, cascade structure TABLE OF CONTENTSAbstract i

5、Keywords iAcknowledgement iv1 Introduction 11.1 System Overview 11.2 Thesis Organization 22 Related Work 33 Integral Images and Features 53.1 Integral Images 53.1.1 Rectangle Integral Image (RII) 53.1.2 Triangle Integral Images 63.2 Features 83.2.1 Extension of Rectangle Features 93.2.2 Triangle Fea

6、tures 104 Learning Strong Classifiers 134.1 Weak Classifier 134.2 Learn a Strong Classifier 144.2.1 Discrete AdaBoost (DA) 144.2.2 Discussion about DA 174.2.3 Modifications to DA 184.2.4 Discussion about Modifications 185 Cascade of Strong Classifiers 205.1 Learning Data 205.2 Learn a Cascaded Class

7、ifier 216 Experimental Results 236.1 Image processing 236.2 Scanning Images 276.3 Experimental Results 276.3.1 Dataset 286.3.2 Selection of Features 296.3.3 System Performance 326.3.4 Error Analyses 347 Conclusions and Future Work 36References vAppendix A: Samples of Detection Results viAppendix B:

8、List of Feature Types Selected in Each Stage viiLIST OF FIGURESFigure 1.1: Flow chart of the face detection system 1Figure 3.1: A rectangle integral image 5Figure 3.2: Triangle Integral Image 1 7Figure 3.3: Triangle Integral Image 2 7Figure 3.4: Face characteristics proposed in 12 8Figure 3.5: Featu

9、re types proposed in 3 8Figure 3.6: Extended rectangle feature types 9Figure 3.7: Calculation of rectangle feature 10Figure 3.8: Compute the sum of pixels in a rectangle area 10Figure 3.9: Type 7: Triangle Feature Type 1 11Figure 3.10: Type 8: Triangle Feature Type 2 11Figure 3.11: Type 9: combinati

10、on of features 12Figure 4.1: Thresholds of a weak classifier 13Figure 4.2: Samples of re-weighting process 17Figure 5.1: Cascaded structure 20Figure 6.1: An image resizing sample 24Figure 6.2: Contrast Stretching 25Figure 6.3: Examples of image processing 26Figure 6.4: Samples of face and non-face i

11、mages 29Figure 6.5: Samples of type 7 and type 8 29Figure 6.6: Total numbers of features of each type 30Figure 6.7: Comparison of training error of using new feature types 31Figure 6.8: FPR of each stage 32Figure 6.9: Comparison of training error of modification to GA 33Figure 6.10: An example of mi

12、sclassification 34LIST OF TABLESTable 4.1: Procedure of Discrete AdaBoost 15Table 5.1: Cascaded classifier learning algorithm 22Table 6.1: Numbers of features of each type 30Table 6.2: Comparison of performance 34Acknowledgement 在清華的兩年中，首先要衷心感謝指導教授張智星老師的指導，無論在做人處事或專業領域上的啟發都讓我獲益良多，使我能順利完成本篇論文；並且感謝口試委

13、員的指導，使本論文更加完善。 另外要感謝MIR實驗室的各位，有和大家的互相砥礪、創意的激發，才使得研究生活不致枯燥乏味。 感謝我的家人這兩年來的支持及關心。 最後要感謝大學的同窗好友jclin，這幾年的生活真的很有意思。1 Introduction Face detection is an important component of a content-based video information retrieval system. There are three main directions for fast face detection researches. The first o

14、ne is to find new useful feature types to decrease the number of classifiers. The second one is to modify existent learning process or introduce a new one to select features. The third one is to decrease the numbers of sub-images to detect to speed up the detection speed. This research focuses on th

15、e first two directions. This research has two main contributions. First, we introduce two novel kinds of integral images and feature types. Second, we observe some problems of Discrete AdaBoost and modify the learning algorithm. This research focuses on detections of upright-frontal faces.1.1 System

16、 OverviewThe flow chart of our system is shown as figure 1.1.Figure 1.1: Flow chart of the face detection system The face detection component is a cascaded structure. The structure is cascaded by strong classifiers. Each strong classifier consists of several weak classifiers. And a weak classifier c

17、onsists of a weight and a feature with thresholds. We can generate many sub-images from an image by various positions and scales. Each strong classifier rejects numbers of sub-images; the rejected sub-images are no longer being processed. Most of those rejected sub-windows are non-face images, and f

18、ew of them are face images. We have to define detection rate (DR) and false positive (FP) first. DR is the ratio of number of face images which are correctly detected to total face number, e.g. 80 faces are detected out of 100 faces, DR = 0.8. FP is a number of non-face images which are detected as

19、face images; the rate of FP to total non-faces is therefore called false positive rate (FPR).1.2 Thesis Organization This thesis is organized as follows: chapter two introduces the related face detection researches; chapter three introduces integral images and features used in our system; chapter fo

20、ur introduces a learning algorithm to select features and train a strong classifier; chapter five introduces a learning process to train a cascaded classifier; chapter six shows the experiments and results; chapter seven talks about the conclusions and future works.2 Related Work In 2001, Viola and

21、Jones introduce a rapid object detection system 12. Their research has three main contributions. First, they introduce integral image which allows fast computation of features. Second, they use AdaBoost 4 to train efficient classifiers. Third, they introduce a cascaded structure which can reject non

22、-face images quickly. Their system can achieve high detection rate with small number of false positives. Later in 2002, Lienhart et al. extend Violas research and their research has three main contributions 3. First, they introduce a novel feature set which is designed for detecting in-plane rotatio

23、n faces. Second, they present analyses among the different boosting algorithms (Discrete, Real and Gentle AdaBoost). Third, they compare the performance between stumps and Regression Tree (CART) and also analyze the effect of sizes of training data. Stan Z. Li and ZhenQiu Zhang introduce a novel lea

24、rning procedure which is called FloatBoost 5. FloatBoost comes from the floating search algorithm. Recall that there are basically three kinds of feature selection methods: Sequential Forward Selection (SFS) which is used in AdaBoost, Sequential Backward Selection (SBS) and Sequential Floating Searc

25、h Method (SFSM) which combines SFS and SBS. SFSM can achieve approximate optimal combination of selection. FloatBoost uses SFSM to select features; the training time is five times longer than AdaBoost. They also introduce a pyramid structure for detecting multiple out-of-plane degree faces. Yong Ma

26、and Xiaoqing Ding introduce Cost Sensitive-AdaBoost (CS-AdaBoost) 6. The weak learner can select more features by using CS-AdaBoost. We propose a similar modification to AdaBoost for selecting more features also. Dong Zhang et al. introduce a face detection framework which uses different kinds of fe

27、atures in early stages and late stages because local features (haar-like features) may not be very useful in late stages 7. Thus they use global feature in late stages. Global feature uses PCA (Principal Component Analysis) features. In the introduction, we bring up the directions of face detection

28、researches. As summary, 12 introduce a framework of fast face detection system, 37 introduce new feature types and 56 modify the training process.3 Integral Images and Features Features used by our system can be computed very fast through integral images. In this chapter, we will introduce integral

29、images, feature types and the way to use integral images to calculate a value from a feature.3.1 Integral Images Integral image is also called Summed Area Table (SAT). It represents a sum of a particular area in an image.3.1.1 Rectangle Integral Image (RII) RII is introduced in 12. The value at posi

30、tion (x, y) in a RII represents the sum of pixels above and left to (x, y) in the original image:,where RII(x, y) is the value of RII at position (x, y) and I(x, y) is the pixel value of the original image at position (x, y).(x, y)Figure 3.1: A rectangle integral imageFor each image, we compute its

31、RII through only one pass of scanning the pixels in an image. In practice, first we compute the cumulative row sum and then add the sum to RII at previous row and the same column to get the sum of pixels above and left:(3-1)(3-2) During computing a RII, we also compute the square sum of the image fo

32、r calculating its variance for contrast stretching; section 6.1 has detailed descriptions.3.1.2 Triangle Integral Images This research introduces two new types of integral images which are called Triangle Integral Images (TIIs). The idea of TIIs comes from the Rotated Summed Area Table (RSAT) in 3. TIIs represent the sums of right triangle areas in an image. TIIs can be