Project Proposal: Robust Face Detection for the facial expression recognition
Jin Hyun, Kim
To make the system, robust face detection module, for the facial expression recognition What is the robust face detection?
– – –
Tolerance of the Background complexity Robustness of the Color variance Brightness & Contrast invariance Real-time Face detection
Inside of the building (ex. The room of the laboratory) On-line Image sequence
80% success rate (10 persons)
Environment of Implementation
– – – –
화소:35만, 해상도:640ⅹ480, frame rate: 30 frames/sec Intel 2.4G Hz CPU, OS: Windows XP
Common Personal Computer
Language: C++ Development tool: Visual C++ 6.0
What is the most adaptive algorithm for the robust face detection in complex environment? How can improve the speed of the system?
효과적인 표정인식을 위해 어떤 정보가 얼굴 검출 의 결과로 나와야 하는가?
Face Detection : A Survey
Erik Hjelmas, Boon Kee Low
Computer Vision and Image Understanding 83, pp236-274 (2001)
Introduction Evolution of Face Detection Research Feature-Based Approach (특징 기반 방법)
Low-Level Analysis Feature Analysis Active Shape Models Linear Subspace Methods Neural Networks Statistical Approaches
Image-Based Approach (이미지 기반 방법)
Applications Summary & Conclusions
Edges, Gray Information, Color, Motion, Generalized Measures
Feature Searching Constellation Analysis Snakes Deformable Template Point Distributed Models
Active Shape Models
Low-level analysis: ambiguous
배경색이 얼굴색과 비슷할 경 우 문제 발생
: 개인의 얼굴 특징들의 상대 적인 위치에 기초한다. : 경험적 정보에 의존
Classical many to one mapping problem
다양한 특징들을 특성화 (characterization)하고 입증 (verification)
: 다양한 얼굴 모델을 사용하 는 유연한 배열 (flexible constellations)
Feature Searching (1)
Determination of prominent facial features
A pair of eyes: commonly applied reference feature
Facial feature extraction algorithm (De Silva)
머리꼭대기(the top of head)를 찾고 눈 영역(eyeplane)을 찾는 순서로… 안경을 쓰고 있고, 머리칼이 이마를 덮을 경우 검출 실패가 있다.
Feature Searching (2)
A system for face and facial feature detection: based on anthropometric measures (Jeng et. al.)
: 후보로 뽑힌 눈에 대해 코, 입 그리고 눈썹 탐색을 한 다. 각 얼굴 특징은 관련 평가함수를 가진다.
Automatic facial feature searching algorithm Based on the motivation of eye movement strategies in the human visual system (HVS)
heuristic information: fixed condition
Grouping facial features in Face-like constellation
Use of statistical shape theory To handle missing feature and problems due to translation, rotation, and scale to a certain extent. 대부분의 검출 실패: 현저한 머리 움직임
Active Shape Models
Actual physical and hence higher-level appearance of features. Three types of the active shape model
Snakes: generic active contour Deformable templates: 안면 특징의 선험성(priori) 를 고려, snake의 향상 PDM (Point distributed model)
Snakes (Active Contour)
Commonly used to locate a head boundary Two main considerations in implementing a snake
The selection of the appropriate energy terms The energy minimization technique
Part of the contour often becomes trapped onto false image features. Snakes are not efficient in extracting nonconvex features due to their tendency to attain minimum curvature.
The concept of snakes + the global information of the eye
A deformable eye template based on its salient features is parameterized using 11 parameters. Working according to the same principle as a snake, the template once initialized near an eye feature will deform itself toward optimal feature boundaries.
Several major considerations
The evolution of a deformable template is sensitive to its initial position The processing time is also very high due to the sequential implementation of the minimization process. The weights of the energy terms are heuristic and difficult to generalize.
Point Distributed Model
PDM is a compact parameterized description of the shape based upon statistics. The advantages of using a face PDM
A compact parameterized description The global characteristic of the model also allows all the features to be located simultaneously and thereby removes the need for feature searching. The occlusion of a particular feature does not pose a severe problem since other features in the model can still contribute to a global optimal solution
The unpredictability of face appearance and environmental conditions. More hostile scenarios: detecting multiple faces with clutter-intensive backgrounds. Pattern recognition problem
Linear Subspace Methods Neural Networks Statistical Approaches
Linear Subspace Methods
Multivariate statistical analysis
PCA (Principle Component Analysis) LDA (Linear discriminant Analysis) FA (Factor Analysis)
PCA (Principle Component Analysis)
Given an ensemble of different face images, the technique first finds the principal components of the distribution of faces, expressed in terms of eigenvectors (of the covariance matrix of the distribution). Each individual face in the face set can then be approximated by a linear combination of the largest eigenvectors, more commonly referred to as eigenfaces, using appropriate weights.
Kullback relative information (Kullback divergence): maximum likelihood face detection
SVM (Support Vector Machine) Bayes’ Decision Rule
How does one count correct detection and false positives? What is the system’s ROC curve? What is the size of the training set and how is training implemented? What is the face?
Biometric identification Video conferencing Indexing of image and video database Intelligent human-computer interface
What is the face? How can we automatically count correct detection and false positives? How can we define the normalized face image in the source image?