Face recognition using Deep Learning

Size: px

Start display at page:

Download "Face recognition using Deep Learning"

Karen Freeman
5 years ago
Views:

1 Face recognition using Deep Learning Master Thesis Author: Serra, Xavier 1 Advisor: Castán, Javier 2 Tutor: Escalera, Sergio 3 1 Master in Artificial Intelligence Barcelona School of Informatics 2 GoldenSpear LLC 3 Department of Mathematics and Computer Science University of Barcelona Barcelona School of Informatics, January 2017 Barcelona School of Informatics Face recognition using Deep Learning 1 / 44

2 Outline 1 Introduction and Goals Introduction Goals 2 How it has been approached Problems Common approaches Proposed approach Datasets 3 Outcome 4 Conclusions Barcelona School of Informatics Face recognition using Deep Learning 2 / 44

3 Outline 1 Introduction and Goals Introduction Goals 2 How it has been approached Problems Common approaches Proposed approach Datasets 3 Outcome 4 Conclusions Barcelona School of Informatics Face recognition using Deep Learning 3 / 44

4 Uses of Face Recognition Face Recognition has drawn plenty of attention It has potential for multiple applications: Biometrical verification Search for a person through cameras Automatically tagging friends Finding similar people... So, what is actually Face Recognition? Barcelona School of Informatics Face recognition using Deep Learning 4 / 44

5 Face Recognition in fiction How has fiction pictured face recognition? Barcelona School of Informatics Face recognition using Deep Learning 5 / 44

6 Actual Face Recognition How does Face Recognition actually work? Eigenfaces Active Appearance Models Support Vector Machines Bayesian models Convolutional Neural Networks... Figure: Example of a CNN Barcelona School of Informatics Face recognition using Deep Learning 6 / 44

7 Goal of this master thesis Developing a face recognition system so that: Keeps a DB of known users Given a new picture, determines the closest match Capable of on-line learning Usable in uncontrolled environments Reasonably fast Barcelona School of Informatics Face recognition using Deep Learning 7 / 44

8 Outline 1 Introduction and Goals Introduction Goals 2 How it has been approached Problems Common approaches Proposed approach Datasets 3 Outcome 4 Conclusions Barcelona School of Informatics Face recognition using Deep Learning 8 / 44

9 Face Recognition Problems Many factors to take into account: Barcelona School of Informatics Face recognition using Deep Learning 9 / 44

10 Face Recognition Problems Many factors to take into account: Light conditions Barcelona School of Informatics Face recognition using Deep Learning 9 / 44

11 Face Recognition Problems Many factors to take into account: Light conditions Expression Barcelona School of Informatics Face recognition using Deep Learning 9 / 44

12 Face Recognition Problems Many factors to take into account: Light conditions Expression Face orientation Barcelona School of Informatics Face recognition using Deep Learning 9 / 44

13 Face Recognition Problems Many factors to take into account: Light conditions Expression Face orientation Age... Barcelona School of Informatics Face recognition using Deep Learning 9 / 44

variability Figure: Intra-class variability

14 Face Recognition Problems Many factors to take into account: Light conditions Expression Face orientation Age... It can be summarized as Intra-class variability Figure: Intra-class variability Barcelona School of Informatics Face recognition using Deep Learning 9 / 44

15 Face Recognition Problems Inter-class similarity is also an issue: Figure: Inter-class similarity Barcelona School of Informatics Face recognition using Deep Learning 10 / 44

16 Common Face Recognition approach Problems of raw images: Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

17 Common Face Recognition approach Problems of raw images: Excessive noise Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

18 Common Face Recognition approach Problems of raw images: Excessive noise Large dimensionality Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

19 Common Face Recognition approach Problems of raw images: Excessive noise Large dimensionality Variability is too high Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

20 Common Face Recognition approach Problems of raw images: Excessive noise Large dimensionality Variability is too high Solution? Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

21 Common Face Recognition approach Problems of raw images: Excessive noise Large dimensionality Variability is too high Solution? Convert input image into a reduced space Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

22 Common Face Recognition approach Problems of raw images: Excessive noise Large dimensionality Variability is too high Solution? Convert input image into a reduced space Feature extraction Manually crafted Automatically found Barcelona School of Informatics Face recognition using Deep Learning 11 / 44

Common approaches Eigenfaces Reduces faces into more compact representations Uses PCA to produce those Set of eigenvectors from the covariance matrix

23 Common approaches Eigenfaces Reduces faces into more compact representations Uses PCA to produce those Set of eigenvectors from the covariance matrix Comparison by linear combination of eigenfaces Figure: Set of eigenfaces Barcelona School of Informatics Face recognition using Deep Learning 12 / 44

Common approaches Active Appearance Models Fits a pre-defined face shape into the image Iteratively improves initial estimation Allows finding sets

24 Common approaches Active Appearance Models Fits a pre-defined face shape into the image Iteratively improves initial estimation Allows finding sets of relevant points Figure: Active Appearance Models fitting a face shape Barcelona School of Informatics Face recognition using Deep Learning 13 / 44

25 Common approaches Support Vector Machines Successful classifier in many problems Finds the hyperplane separating two problems Can be used to determine if two images belong to same person Figure: Application of Support Vector Machines Barcelona School of Informatics Face recognition using Deep Learning 14 / 44

26 Common approaches Bayesian models Models each facial feature as x = µ + ɛ It corresponds to inter-class and intra-class variability Based on the full joint distribution of face image pairs r(x 1, x 2 ) = log P(x 1, x 2 H I ) P(x 1, x 2 H E ) Barcelona School of Informatics Face recognition using Deep Learning 15 / 44

Common approaches Convolutional Neural Network It is a type of Artificial Neural Network Works by finding increasingly abstract features Takes into account spatial relation High requirements

27 Common approaches Convolutional Neural Network It is a type of Artificial Neural Network Works by finding increasingly abstract features Takes into account spatial relation High requirements in time and data Currently providing state of art results in many CV problems Figure: Convolutional Neural Network Barcelona School of Informatics Face recognition using Deep Learning 16 / 44

28 Proposed approach The proposed approach consists of 4 steps: Barcelona School of Informatics Face recognition using Deep Learning 17 / 44

29 Proposed approach The proposed approach consists of 4 steps: Step 1: Locating the main face in the image Barcelona School of Informatics Face recognition using Deep Learning 17 / 44

30 Proposed approach The proposed approach consists of 4 steps: Step 1: Locating the main face in the image Step 2: Frontalizing the found face Barcelona School of Informatics Face recognition using Deep Learning 17 / 44

31 Proposed approach The proposed approach consists of 4 steps: Step 1: Locating the main face in the image Step 2: Frontalizing the found face Step 3: Extracting features using a CNN Barcelona School of Informatics Face recognition using Deep Learning 17 / 44

32 Proposed approach The proposed approach consists of 4 steps: Step 1: Locating the main face in the image Step 2: Frontalizing the found face Step 3: Extracting features using a CNN Step 4: Performing comparison with stored ones Barcelona School of Informatics Face recognition using Deep Learning 17 / 44

33 Step 1: Locate the face Goal: Look for the bounding box of the most likely face Figure: Locating the face Benefit: Prevent erroneously located faces in next step Barcelona School of Informatics Face recognition using Deep Learning 18 / 44

Step 1: Locate the face Procedure: Using a region based Convolutional Neural Networks (Faster RCNN [RHGS15]) Set of possible face locations is produced Most

34 Step 1: Locate the face Procedure: Using a region based Convolutional Neural Networks (Faster RCNN [RHGS15]) Set of possible face locations is produced Most promising face is kept: distance to center + confidence Figure: Selecting most likely face Barcelona School of Informatics Face recognition using Deep Learning 19 / 44

35 Step 2: Frontalize the face Goal: Frontalize the face so that it is looking at the camera Figure: Frontalizing the found face Benefits: Eliminate background noise + Equally placed faces Barcelona School of Informatics Face recognition using Deep Learning 20 / 44

Step 2: Frontalize the face Procedure: 1 Locate a set of 46 fiducial points 2 Consider the same points in a 3D pre-defined model 3 Generate a projection matrix to map from 2D input to the

36 Step 2: Frontalize the face Procedure: 1 Locate a set of 46 fiducial points 2 Consider the same points in a 3D pre-defined model 3 Generate a projection matrix to map from 2D input to the 3D reference 4 (Apply vertical similarity to fill in empty spots) Discarded Figure: Frontalization process [HHPE15] Barcelona School of Informatics Face recognition using Deep Learning 21 / 44

37 Step 2: Frontalize the face Not working perfectly: Figure: Examples of successful and unsuccessful frontalizations Barcelona School of Informatics Face recognition using Deep Learning 22 / 44

38 Step 3: Extract relevant features Goal: Automatically extract a set of relevant features from the face Benefits: More efficient comparison + Reduction in variability Barcelona School of Informatics Face recognition using Deep Learning 23 / 44

Step 3: Extract relevant features Goal: Automatically extract a set of relevant features from the face Benefits: More efficient comparison + Reduction in variability Procedure: A CNN has been used to

39 Step 3: Extract relevant features Goal: Automatically extract a set of relevant features from the face Benefits: More efficient comparison + Reduction in variability Procedure: A CNN has been used to process each image Each image is compressed into a reduced representation A feature vector of 4096 features is generated Based on Facebook s DeepFace method [TYRW14] Figure: CNN architecture used Barcelona School of Informatics Face recognition using Deep Learning 23 / 44

40 Step 4: Compare them Goal: Perform the comparison with DB to look for a match Procedure: 1 Given a generated feature vector g 2 Iterates over all people in the DB 3 Each person has N relevant feature vectors F = f 1, f 2,...f N 4 Distance comparison is performed between g and each f i F 5 Various selection measures considered: minimum, mean, etc. Barcelona School of Informatics Face recognition using Deep Learning 24 / 44

41 Datasets used Three datasets considered: Casia dataset: 495,000 pictures / 10,500 people CACD dataset: 160,000 pictures / 2,000 people FaceScrub: 100,000 pictures / 500 people Training: 500,000 pictures / 9,351 people Testing: 100,000 pictures / 1,671 people Additionally, to use as a benchmark: Labeled Faces in the Wild: 13,000 pictures / 5,700 people Barcelona School of Informatics Face recognition using Deep Learning 25 / 44

42 Datasets used Generated datasets From training dataset, we generated two extra: Barcelona School of Informatics Face recognition using Deep Learning 26 / 44

43 Datasets used Generated datasets From training dataset, we generated two extra: Augmented: Using data augmentation Randomly modifying light intensity Other data augmentations made not much sense rotation, scaling, etc. 1M instances Barcelona School of Informatics Face recognition using Deep Learning 26 / 44

44 Datasets used Generated datasets From training dataset, we generated two extra: Augmented: Using data augmentation Randomly modifying light intensity Other data augmentations made not much sense rotation, scaling, etc. 1M instances Grayscale: Convert previous dataset to grayscale Aims to make the problem easier for CNN Both training and testing sets converted CNN modified accordingly Barcelona School of Informatics Face recognition using Deep Learning 26 / 44

45 Outline 1 Introduction and Goals Introduction Goals 2 How it has been approached Problems Common approaches Proposed approach Datasets 3 Outcome 4 Conclusions Barcelona School of Informatics Face recognition using Deep Learning 27 / 44

46 How to evaluate their performance? Face Recognition systems can be evaluated according to: Face Verification Face Recognition Barcelona School of Informatics Face recognition using Deep Learning 28 / 44

47 How to evaluate their performance? Face Recognition systems can be evaluated according to: Face Verification Face Recognition Both intrinsically related... Barcelona School of Informatics Face recognition using Deep Learning 28 / 44

48 How to evaluate their performance? Face Recognition systems can be evaluated according to: Face Verification Face Recognition Both intrinsically related but differently evaluated Barcelona School of Informatics Face recognition using Deep Learning 28 / 44

49 Face Verification Description Goal: Determining whether two pictures belong to same person: Needed on most Face Recognition systems Performance not directly related with Face Recognition step Commonly used as benchmark to compare methods The Labeled Faces in the Wild dataset has been used 2000 training pairs / 1000 test pairs Allowed for hyperparameter tuning Barcelona School of Informatics Face recognition using Deep Learning 29 / 44

50 Face Verification Examples Figure: Example on test pairs Barcelona School of Informatics Face recognition using Deep Learning 30 / 44

51 Face Verification Procedure Comparison performed using Euclidean and Taxicab distances Weighted variations considered but discarded due to bad results Training consists in: 1 Obtain distance between all train pairs 2 Find the optimal threshold placement to separate classes Barcelona School of Informatics Face recognition using Deep Learning 31 / 44

52 Face Verification Procedure Comparison performed using Euclidean and Taxicab distances Weighted variations considered but discarded due to bad results Training consists in: 1 Obtain distance between all train pairs 2 Find the optimal threshold placement to separate classes Figure: Example best case scenario Figure: Example more difficult scenario Barcelona School of Informatics Face recognition using Deep Learning 31 / 44

53 Face Verification Results a Method Accuracy Ours Joint Bayesian Tom-vs-Pete High-dim LBP TL Joint Bayesian FaceNet DeepFace Human performance Reasons Too few training data Further need for parameter tuning Improve distance metric Table: Results state of art methods Barcelona School of Informatics Face recognition using Deep Learning 32 / 44

54 Results Figure: Accuracy according to dataset Figure: Accuracy according to distance Barcelona School of Informatics Face recognition using Deep Learning 33 / 44

55 Face Recognition Description Goal: Determining who the person is: Select among a set of people in a DB Person-wise comparison Face Verification Closest match is selected Need to determine if there is a match at all Seemingly more difficult than Face Verification empirical results prove it may not be so Barcelona School of Informatics Face recognition using Deep Learning 34 / 44

56 Face Recognition Procedure Reminder: comparing feature vector f with all people in DB Each person has N feature vectors F = f 1, f 2,...f N Comparison strategies: Barcelona School of Informatics Face recognition using Deep Learning 35 / 44

57 Face Recognition Procedure Reminder: comparing feature vector f with all people in DB Each person has N feature vectors F = f 1, f 2,...f N Comparison strategies: 1 Distance to closest feature vector in F Barcelona School of Informatics Face recognition using Deep Learning 35 / 44

58 Face Recognition Procedure Reminder: comparing feature vector f with all people in DB Each person has N feature vectors F = f 1, f 2,...f N Comparison strategies: 1 Distance to closest feature vector in F 2 Mean distance to all f i F Barcelona School of Informatics Face recognition using Deep Learning 35 / 44

59 Face Recognition Procedure Reminder: comparing feature vector f with all people in DB Each person has N feature vectors F = f 1, f 2,...f N Comparison strategies: 1 Distance to closest feature vector in F 2 Mean distance to all f i F 3 Product of 1 and 2 Barcelona School of Informatics Face recognition using Deep Learning 35 / 44

60 Face Recognition Procedure Reminder: comparing feature vector f with all people in DB Each person has N feature vectors F = f 1, f 2,...f N Comparison strategies: 1 Distance to closest feature vector in F 2 Mean distance to all f i F 3 Product of 1 and 2 4 Product of distance to furthest feature vector in f and 3 Barcelona School of Informatics Face recognition using Deep Learning 35 / 44

61 Face Recognition Procedure Reminder: comparing feature vector f with all people in DB Each person has N feature vectors F = f 1, f 2,...f N Comparison strategies: 1 Distance to closest feature vector in F 2 Mean distance to all f i F 3 Product of 1 and 2 4 Product of distance to furthest feature vector in f and 3 The smallest distance is chosen as a match Barcelona School of Informatics Face recognition using Deep Learning 35 / 44

62 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

63 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

64 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

65 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it 3 If M higher than T 3, discard it (extreme outlier) Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

66 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it 3 If M higher than T 3, discard it (extreme outlier) 4 Select the feature vectors - F O - far from mean (outliers) Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

67 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it 3 If M higher than T 3, discard it (extreme outlier) 4 Select the feature vectors - F O - far from mean (outliers) 5 Face Verification between f and all f i F O Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

68 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it 3 If M higher than T 3, discard it (extreme outlier) 4 Select the feature vectors - F O - far from mean (outliers) 5 Face Verification between f and all f i F O 6 If less than half matches, keep it (rare enough case) Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

69 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it 3 If M higher than T 3, discard it (extreme outlier) 4 Select the feature vectors - F O - far from mean (outliers) 5 Face Verification between f and all f i F O 6 If less than half matches, keep it (rare enough case) 7 If more than T 4 feature vector stored, discard closest to mean Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

70 Face Recognition Keeping Procedure Each new feature vector f may be kept into the system: 1 If less than T 1 feature vectors stored, keep it 2 If distance M between f and mean of F less than T 2, discard it 3 If M higher than T 3, discard it (extreme outlier) 4 Select the feature vectors - F O - far from mean (outliers) 5 Face Verification between f and all f i F O 6 If less than half matches, keep it (rare enough case) 7 If more than T 4 feature vector stored, discard closest to mean T 1, T 2, T 3 and T 4 are hyperparameters Barcelona School of Informatics Face recognition using Deep Learning 36 / 44

71 Face Recognition Dataset Self-generated dataset, from training dataset: 100 people (50 females / 50 males) 30 training images each 50 training images each Manually cleaned Barcelona School of Informatics Face recognition using Deep Learning 37 / 44

72 Face Recognition Results Figure: Accuracy according to number kept images Figure: Accuracy according to comparison strategy Barcelona School of Informatics Face recognition using Deep Learning 38 / 44

comparison strategy A 95% of accuracy was reached

73 Face Recognition Results Figure: Accuracy according to number kept images Figure: Accuracy according to comparison strategy A 95% of accuracy was reached Barcelona School of Informatics Face recognition using Deep Learning 38 / 44

74 Outline 1 Introduction and Goals Introduction Goals 2 How it has been approached Problems Common approaches Proposed approach Datasets 3 Outcome 4 Conclusions Barcelona School of Informatics Face recognition using Deep Learning 39 / 44

75 To conclude... We have developed a functional Face Recognition System using CNNs Works in uncontrolled environment, capable of on-line learning Compared with state of art methods, it underperforms in Face Verification Quality results achieved in Face Recognition Exhaustive tests performed reliable results Barcelona School of Informatics Face recognition using Deep Learning 40 / 44

76 ... or not! Future work lines: Improve CNN performance: More data Better parameter tuning Test more comparison metrics: Further try thresholding strategies Different weights Enhance matching capabilities: Use more complex strategies apart from min, mean, etc. Modify on-line learning mechanism Consider other alternatives for feature extraction: Other existing approaches Develop one on our own Barcelona School of Informatics Face recognition using Deep Learning 41 / 44

77 References I Tal Hassner, Shai Harel, Eran Paz, and Roee Enbar, Effective face frontalization in unconstrained images, IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), June Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, Advances in Neural Information Processing Systems 28 (C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, eds.), Curran Associates, Inc., 2015, pp Yaniv Taigman, Ming Yang, Marc Aurelio Ranzato, and Lior Wolf, Deepface: Closing the gap to human-level performance in face verification, Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition (Washington, DC, USA), CVPR 14, IEEE Computer Society, 2014, pp Barcelona School of Informatics Face recognition using Deep Learning 42 / 44

78 Questions? Any Question? Barcelona School of Informatics Face recognition using Deep Learning 43 / 44

79 Thank you! Thank you for your attention! Barcelona School of Informatics Face recognition using Deep Learning 44 / 44

Lecture 1: Machine Learning Basics

1/69 Lecture 1: Machine Learning Basics Ali Harakeh University of Waterloo WAVE Lab ali.harakeh@uwaterloo.ca May 1, 2017 2/69 Overview 1 Learning Algorithms 2 Capacity, Overfitting, and Underfitting 3