Discrete area filters in accurate detection of faces and facial features☆
Introduction
Accurate face and facial feature detection plays a major role in image processing techniques. In order to perform reliable face recognition, significant parts of the faces must be precisely localized, for example eye centers. Lack of accuracy highly decreases efficiency of the identification. Simple holistic face detection algorithms, for example the one used in popular Viola and Jones method, are trained in such fashion, that the eye centers are located in the specific locations of the bounding box. In practice the position of the detected faces isn't very precise and it gives only regional information. In order to perform reliable face recognition more sophisticated algorithms should be used.
Many methods cope with this problem, but they usually suffer from a subset of several disadvantages — model isn't flexible enough to cover all the variability of the human face (AAM, ASM), different treatment is needed for each part of the face [1], or only localisation of one face is possible.
There is also a possibility of refining face/facial feature position already given by the bounding box detector. Such algorithms often use content versus context approach [1] or detailed knowledge-based rules. Such top-down approach (from general information to details) often ensures high speed of the localization, but on the other hand, holistic methods tend to reject partially occluded faces. Some of the algorithms are trained using also such difficult examples, but the final detector tends to have rather lower accuracy on the non-occluded faces.
This article presents new bottom-up (from details to the general information) algorithm for detecting facial features (eyes, nose, mouth) and faces. It introduces new extraction and classifications methods for detecting characteristic face points (such as eye corner and mouth corner) — discrete area filters (DAF) and modified positive-to-negative linear discriminant analysis (mLDA). Points belonging to the specific categories are merged into face parts and implicitly to faces by the voting process combined with agglomerative clustering and the reference graph. Article proves that such an approach yields accurate detection comparing to manual marking, high detection rates and allows the detection of partially occluded faces.
Paper is divided as follows:
- •
Section 2 presents problem formulation and current state-of-the-art in the face and facial feature detection and localization.
- •
Section 3 provides basic information about each step of the algorithm, including discrete area filters and modified linear discriminant analysis.
- •
Section 4 describes performed experiments along with their interpretation.
- •
Section 5 concludes the paper.
Section snippets
Problem formulation and related work
In this paper fiducial point detection stands for defining number and positions of all the points of specific category, for example inner eye corners or nose tips. Facial feature detection or face part detection stands for defining number and positions of all complex parts of the face such as nose, mouth or eyes. Face detection stands for defining the number and position of all the faces present in the digital image. This can be achieved in several ways, for example by defining a bounding box,
Algorithm description
Face detection algorithm is designed in a bottom-up fashion. After defining prior points of interest — edge points, each point is classified into one of the defined categories (Fig. 1). Each point votes for the center of the face feature which is a part of, basing on the vectors assigned to each class, similarly to the generalized Hough transform (for example eye corner votes for the eye center). After detecting independently eyes, noses and mouths, we compare responses to the reference graph
Algorithm training
Algorithm's parameters were trained in a supervised fashion. Note, that each fiducial point/non-fiducial point classifier is trained separately, resulting in 9 detectors. We chose 1980 face images with marked face points (positive examples) and a set of 10,000 non-face images (negative examples). We also defined a set of possible discrete area filters (Appendix A). To assure high variability of the provided faces, positive examples were gathered from various databases:
- •
Banca — subset used in a
Conclusion
This paper presents a novel algorithm for face and facial feature detection. Proposed extraction scheme based on new discrete area filters (DAF) and modification of the linear discriminant analysis proves to be robust and fast. Accurate results, comparable to the manual markings, and on-line detection make this method competitive to the current state-of-the-art. Algorithms introduce no bounds to the number of the detected faces and prove to introduce very low number of false acceptance, for
Jacek Naruniec received his M.Sc. and Ph.D. at Warsaw University of Technology in 2006 and 2010, respectively. His work is mainly concentrated on semantic image analysis, especially human faces. He is currently working as an assistant professor in the Institute of Radioelectronics at Warsaw University of Technology and as a researcher at the Center of Cosmic Research in Polish Academy of Sciences in Warsaw. His courses refer mainly to semantic object analysis and object programming.
References (56)
- et al.
Robust real-time eye detection and tracking under variable lighting conditions and various face orientations
Comput. Vis. Image Underst.
(2005) - et al.
Active shape models — their training and application
Comput. Vis. Image Underst.
(1995) Generalizing the Hough transform to detect arbitrary shapes
Pattern Recogn.
(1981)- et al.
Features versus context: an approach for precise and detailed detection and delineation of faces and facial features
IEEE Trans. Pattern Anal. Mach. Intell.
(nov. 2010) - et al.
Man–machine interaction in human face identification
Bell Syst. Tech.
(feb 1972) - et al.
The representation and matching of pictorial structures
IEEE Trans. Comput.
(1973) - et al.
Independent component analysis and support vector machine for face feature extraction
- et al.
Multi-attribute robust facial feature localization
- et al.
Discriminative regions for human face detection
- et al.
Feature-based affine-invariant localization of faces
IEEE Trans. Pattern Anal. Mach. Intell.
(2005)
Active appearance models
Lect. Notes Comput. Sci
Eigenfaces vs. Fisherfaces: Recognition Using Class Specific Linear Projection
Component-based lda face descriptor for image retrieval
Scale invariant face detection method using higher-order local autocorrelation features extracted from log-polar image
Eye and face detection using SVM
Facial feature detection with optimal pixel reduction SVMs
Neural network-based face detection
IEEE Trans. Pattern Anal. Mach. Intell.
Locating facial features in image sequences using neural networks
A connexionist approach for robust and precise facial feature detection in complex scenes
Robust real-time face detection
Int. J. Comput. Vis.
Fast and robust classification using asymmetric AdaBoost and a detector cascade
Empirical Analysis of Detection Cascades of Boosted Classifiers for Rapid Object Detection
Classifying facial actions
IEEE Trans. Pattern Anal. Mach. Intell.
Distortion invariant object recognition in the dynamic link architecture
IEEE Trans. Comput.
Face recognition by elastic bunch graph matching
IEEE Trans. Pattern Anal. Mach. Intell.
Automated Facial Feature Detection From Portrait and Range Images
Fully automatic facial feature point detection using Gabor feature based boosted classifiers
A comparison of Gabor filter methods for automatic detection of facial landmarks
Cited by (0)
Jacek Naruniec received his M.Sc. and Ph.D. at Warsaw University of Technology in 2006 and 2010, respectively. His work is mainly concentrated on semantic image analysis, especially human faces. He is currently working as an assistant professor in the Institute of Radioelectronics at Warsaw University of Technology and as a researcher at the Center of Cosmic Research in Polish Academy of Sciences in Warsaw. His courses refer mainly to semantic object analysis and object programming.
- ☆
This paper has been recommended for acceptance by Stefanos Zafeiriou.