Fully Automatic Facial Feature Point Detection Using Gabor Feature Based Boosted Classifiers
- 1. JAMES COOK AUSTRALIA INSTITUTE OF HIGHER LEARNING IN SINGAPORE HEALTH DIAGNOSTIC BY ANALYSING FACE IMAGES USING MOBILE DEVICES Instructor : Dr. Insu Song Student : Ho Thi Hoang Yen Email:
- 2. INTRODUCTION
- 3. INTRODUCTION A robust, highly accurate method for detecting 20 facial points in images of expressionless faces
- 4. INTRODUCTION Facial feature points are generally referred to as facial salient points such as : 1. the corners of the eyes (ABFG) 2. corners of the eyebrows (ED) 3. corners and outer mid points of the lips (IKJL) 4. corners of the nostrils (H) 5. tip of the nose (N) 6. the tip of the chin (M)
- 5. How do they localization facial feature points ? Currently, localization of facial points is usually carried out by manually labeling the required set of points. Previous methods categories: 1. Texture-based Methods: modeling local texture around a given feature point. 2. Shape-based Methods: regard all facial feature points as a shape. => None of them detects all 20 facial feature points and None has high accuracy. INTRODUCTION
- 6. INTRODUCTION Face detection using Haar feature based GentleBoost classifier feature extraction based on Gabor filtering d) feature selection and classification using GentleBoost classifier, output of the system compared to the face drawing with facial landmark points we aim to detect
- 7. The Method Consists Of 4 Steps: 1. Face Detection 2. Region Of Interest (ROI) Detection 3. Feature Extraction 4. Feature Classification. METHODOLOGY
- 8. METHODOLOGY 1. FACE DETECTION => To build a system capable of automatically labeling facial feature points in a face image, it is first necessary to localize the face in the image. => Using robust real-time face detection of Viola-Jones (replace adaboost with gentleboost) Results : - Training data : 5000 faces and millions of non-face patches from the web - Detecting test : 422 images from the Cohn-Kanade database => 100% detection rate
- 9. The Method Consists Of 4 Steps: 1. Face Detection 2. Region Of Interest (ROI) Detection 3. Feature Extraction 4. Feature Classification. METHODOLOGY
- 10. METHODOLOGY 2. Region Of Interest (ROI) Detection Divide into 2 part : iris part and mouth part. 1. Divide the face region horizontally into two parts 2. The upper face region is again divided into two halves in a vertical direction so that each eye can be analyzed separately.
- 11. METHODOLOGY 2. Region Of Interest (ROI) Detection ANALYSIS THE IRISES => Locate the irises (by vertical and horizontal histograms: comparing rows & columns pixels) => x & y coordinate. => Rotate the images if necessary
- 12. METHODOLOGY 2. Region Of Interest (ROI) Detection ANALYSIS THE MOUTH With those distance, we can locate the medial point of mouth. Test on Cohn-Kanade database: 99% correct rate. (2 cases fail) Use Iris and medial points to divide the face into 20 regions for further calculations.
- 13. The Method Consists Of 4 Steps: 1. Face Detection 2. Region Of Interest (ROI) Detection 3. Feature Extraction 4. Feature Classification. METHODOLOGY
- 14. METHODOLOGY 3. Feature Extraction Divided the face into 20 ROIs, each ROIs corresponds to a facial point to be detected. Gabor wavelet method is used to do this extraction. A slide window 13x13 size slid to the ROIs and get the facial point for extraction. Feature vector for each facial point is extracted from the 13×13 pixels image patch centered on that point f0 is the central frequency of a sinusoidal plane wave, θ is the anti-clock wise rotation of the Gaussian and the plane wave, and α and β are the parameters for scaling two axis of the elliptic Gaussian envelope.
- 15. The Method Consists Of 4 Steps: 1. Face Detection 2. Region Of Interest (ROI) Detection 3. Feature Extraction 4. Feature Classification. METHODOLOGY
- 16. METHODOLOGY 4. Feature Classification. - GentleBoost feature templates are learned using a representative set of positive and negative examples. - The size of a example point (feature vector) is 8281 size matrix. - In contrast to Adaboost, Gentleboost uses real valued features Faster . Perform better for objects detection problems. Numerically robust . Outperform other boosting algorithms (PCA , FLD , LFA).
- 17. METHODOLOGY 4. Feature Classification. Positive and negative examples for training point The big white square on the inner corner of the eye represents 9 positive examples. Around that square are 8 negative examples randomly chosen near the positive examples. Another 8 negative examples are randomly chosen from the rest of the region.
- 18. METHODOLOGY 4. Feature Classification. There is a 25×8281 size matrix representing training data for each ROI for each training image. (9 pos + 16 neg) In the testing phase, each ROI is filtered first by the same set of Gabor filters used in the training phase (in total, 48 Gabor filters are used). For each position of the sliding window, GentleBoost classifier outputs a response depicting the similarity between the 49- dimensional representation of the sliding window compared to the learned feature point model.
- 19. RESULT Training set The facial feature detection method was trained and tested on the Cohn-Kanade database, which consists of approximately 2000 gray-scale image sequences in nearly frontal view from over 200 subjects, male and female, being 18 to 50 years old. => but authors used only first 300 frames for training and testing.
- 21. RESULT
- 22. CONCLUSION • A robust, highly accurate method for fully automatic detection of 20 facial feature points in images of expressionless faces • Using Gabor feature based boosted classifiers. • When tested on images from the Cohn-Kanade database, with possible in-plane head rotations and recorded under various illumination conditions, the method has achieved average recognition rates of 93%.
- 23. SWOT • STRENGTH ? Robust, high accuracy, automatically detection of 20 feature points, not sensitive with head-pose & illumination • WEAKNESS? Not very effective with expressional faces ( 93% rate based on 300 public expressionless samples of CK database ) => authors didn’t guarantee that the method’s performance reported below will remain the same • OPPORTUNITY ? Can be very useful for face detection related programs • THREAT ? Results with public images (not from Cohn-Kanade) may not as effected.
- 24. OPINION This paper has described very clearly the algorithm for face detection. It is valuable for all kinds of researches related to face, to the system for interacting between human & computer , or the face recognition and to FACE ANALYSIS FOR HEALTH PURPOSE.
Editor's Notes
- #2: Good morning every body, My name is Hoang Yen, under guidance of Dr.Insu, I’m researching on the topic health diagnostic by face analysing.
- #3: In order to complete this task, I have to solve the problem : Facial features extraction, to teach the computer how a human face looks like. Among so many methods for facial features extraction, I found a very key paper which is really valuable.
- #4: It’s a paper that is presented in 2005 : Fully Automatic Facial Feature Point Detection using Gabor Feature Based Boosted Classifiers. We propose here a robust, highly accurate method for detecting 20 facial points in images of expressionless faces with possible in-plane rigid head rotations, recorded under various illumination conditions
- #5: At first, I want to talk a little bit about Facial Feature Points. In order to teach the computer the appearance of a human face, we have to convert face image into facical feature points. This is a very core task for all kind of program/ topic/ research those are related to face detection / recognition
- #6: Texture base : dac trung dien mao - less complex – less accuracy – use the illumination of pixel to extract attributes. Shape base : dac trung hinh hoc - use the images of eyes, mouth, eyebrown,.. To localize some point in the face : iris , conner of mouth, … Hybrid : mix of the two.
- #7: Face detection using Haar feature based GentleBoost classifier; (b) ROI extraction, (c) feature extraction based on Gabor filtering, (d) feature selection and classification using GentleBoost classifier, (e) output of the system compared to the face drawing with facial landmark points we aim to detect
- #8: The method consists of 4 steps (Fig. 1): Face Detection, Region Of Interest (ROI) Detection, Feature Extraction, and Feature Classification.
- #9: Gentleboost is an improved version of adaboost. The adapted version of the Viola-Jones face detector that we employ uses GentleBoost instead of AdaBoost. It also refines the originally proposed feature selection by finding the best performing single-feature classifier from a new set of filters generated by shifting and scaling the chosen filter by two pixels in each direction, as well as composite filters made by reflecting each shifted and scaled feature horizontally about the center and superimposing it on the original.
- #10: The method consists of 4 steps (Fig. 1): Face Detection, Region Of Interest (ROI) Detection, Feature Extraction, and Feature Classification.
- #14: The method consists of 4 steps (Fig. 1): Face Detection, Region Of Interest (ROI) Detection, Feature Extraction, and Feature Classification.
- #15: f0 is the central frequency of a sinusoidal plane wave, θ is the anti-clock wise rotation of the Gaussian and the plane wave, and α and β are the parameters for scaling two axis of the elliptic Gaussian envelope.
- #16: The method consists of 4 steps (Fig. 1): Face Detection, Region Of Interest (ROI) Detection, Feature Extraction, and Feature Classification.
- #17: Gabor approach for local feature extraction outperformed PCA (Principal Component Analysis), FLD (Fisher’s Linear Discriminant) and LFA (Local Feature Analysis)
- #24: Gabor approach for local feature extraction outperformed PCA (Principal Component Analysis), FLD (Fisher’s Linear Discriminant) and LFA (Local Feature Analysis)
- #25: Gabor approach for local feature extraction outperformed PCA (Principal Component Analysis), FLD (Fisher’s Linear Discriminant) and LFA (Local Feature Analysis)