ALGORITHM OF PREPARATION OF THE TRAINING SAMPLE USING 3D-FACE MODELING
Abstract
The algorithm of preparation and sampling for training of the multiclass qualifier of support vector machines (SVM) is provided. The described approach based on the modeling of possible changes of the face features of recognized person. Additional features like perspectives of shooting, conditions of lighting, tilt angles were introduced to get improved identification results. These synthetic generated changes have some impact on the classifier learning expanding the range of possible variations of the initial image. The classifier learned with such extended example is ready to recognize unknown objects better. The age, emotional looks, turns of the head, various conditions of lighting, noise, and also some combinations of the listed parameters are chosen as the key considered parameters for modeling. The third-party software ‘FaceGen’ allowing to model up to 150 parameters and available in a demoversion for free downloading is used for 3D-modeling.
The SVM classifier was chosen to test the impact of the introduced modifications of training sample. The preparation and preliminary processing of images contains the following constituents like detection and localization of area of the person on the image, assessment of an angle of rotation and an inclination, extension of the range of brightness of pixels and an equalization of the histogram to smooth the brightness and contrast characteristics of the processed images, scaling of the localized and processed area of the person, creation of a vector of features of the scaled and processed image of the person by a Principal component analysis (algorithm NIPALS), training of the multiclass SVM-classifier.
The provided algorithm of expansion of the training selection is oriented to be used in practice and allows to expand using 3D-models the processed range of 2D – photographs of persons that positively affects results of identification in system of face recognition. This approach allows to compensate in a certain measure a lack of initial images for identification improvement of quality.
About the Authors
D. I. Samal
Belarusian State University of Informatics and Radioelectronics
Belarus
Belarus
Associate Professor, PhD in Engineering, Head of Department
I. I. Frolov
Belarusian State University of Informatics and Radioelectronics
Belarus
Associate Professor, PhD in Engineering
Belarus
Associate Professor, PhD in Engineering
References
1. Starovojtov V. V. Rasshirenie obuchajushhego mnozhestva dlja nastrojki biometricheskih sistem raspoznavanija / B. B. Starovojtov, Nikodimov D. Ju. // Obrabotka informacii i upravlenie v chrezvychajnyh i jekstremal’nyh situacijah: materialy 4-j mezhdunarodnoj konferencii. – Minsk, 2004. – S. 204–209.
2. Yanushkevich S. N. Biometric-Based Decision Support Assistance in Physical Access Control Systems / S. Chague, B. Droit, O. Boulanov, S. N. Yanushkevich, V. P. Shmerko, A. Stoica // Bio-inspired Learning and Intelligent Systems for Security: Proceedings of ECSIS Symposium, Aug. 4–6, 2008, Edinburgh, UK. – Edinburg, 2008. – P. 11–16.
3. Gonsales R. Cifrovaja obrabotka izobrazhenij / R. Gonsales, R. Vuds. – Moskva: Tehnosfera. – 2005. – 1072 s.
4. Hajkin S. Nejronnye seti: polnyj kurs / S. Hajkin. – Moskva: Vil’jams. – 2006. – 1104 s.
Review
For citations:
Samal D.I., Frolov I.I. ALGORITHM OF PREPARATION OF THE TRAINING SAMPLE USING 3D-FACE MODELING. «System analysis and applied information science». 2016;(4):17-23. (In Russ.)
Views:
1016
Email this article 