Method for reducing computational complexity of face detecting teaching procedure based on Viola-Jones method

Аuthors

Kozirevskiy V. K.^*, Veselov A. I.^**

Saint Petersburg State University of Aerospace Instrumentation, 67, Bolshaya Morskaya str., Saint Petersburg, 190000, Russia

*e-mail: vadikko2@mail.ru
**e-mail: felix@vu.spb.ru

Abstract

The article suggests a method for computational complexity reducing of Viola-Jones face detecting teaching procedure. The reference detector teaching procedure is based on AdaBoost algorithm, representing an iterative procedure of algorithmic composition construction (strong classifier). Each element (weak classifier) of the composition analyzes only one feature of an image. In Viola-Jones method feature is the result of image convolution by some pre-defined filter. A set of possible filters is developed in advance and the best feature in terms of strong classifier’s classification errors decreasing is selected at each learning iteration of AdaBoost. The proposed method is a modification of AdaBoost algorithm where complexity reduction is achieved by adaptive selection of the features to be analyzed. The idea of the modified method is based on the observation that at each iteration of the learning procedure the reference algorithm selects weak classifiers, correcting errors of the previously selected weak classifiers. It means that the weak classifiers with similar performance are unlikely to be selected at the adjacent iterations. Thus the space of the tested weak classifiers can be reduced. The proposed method includes two modifications of the reference algorithm. The first modification consists in a feature pre-analysis operation which is performed to estimate the correlation between the responses of different filters. This stage is performed before the reference AdaBoost algorithm is started. In the process of teaching, only the features with low correlation can be analyzed at the adjacent iteration, thus reducing the number of weak classifiers to be evaluated for boosting the final strong classifier.

Practical relevance: the proposed algorithm reduces the computational complexity of Viola-Jones face detector teaching, which is an open issue in computer-vision-related systems.

Keywords:

computer vision, face detection, Viola-Jones algorithm, boosting, AdaBoost, FFS, complexity reduction

References

1. Viola P., Jones M. Robust Real-time Object Detection, International Journal of Computer Vision, 2004, vol. 57, no. 2, pp. 137-154.

2. V’yugin V.V. Matematicheskie osnovy teorii mashinnogo obucheniya (Mathematical Foundations of Machine Learning Theory), Moscow, MTsNMO, 2013, 387 p.

3. Mestetskii L.M. Matematicheskie metody raspoznavaniya obrazov (Mathematical methods of pattern recognition), Moscow, MGU, VMiK, 2002–2004, pp. 42–44.

4. Hastie T., Tibshirani R., Friedman J. The elements of statistical learning, New York, Springer, 2001, 533 p.

5. Viola P., Jones M. Rapid object detection using a boosted cascade of simple features. Accepted conference on computer vision and pattern recognition, Cambridge, 2001, available at: http://www.cs.cmu.edu/~srini/15-829/readings/ViJo01.pdf

6. Freund Y., Schapire R. A decision-theoretic generalization of on-line learning and an application to boosting, Journal of Computer and System Sciences, 1997, pp. 119-139.

7. Wu J, Brubaker S., Mullin M., Rehg J. Fast Asymmetric Learning for Cascade Face Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2008, vol. 30, no. 3 pp. 369-382.

8. CBCL Face Database #1. MIT Center For Biological and Computation Learning, 2000, available at: http://cbcl.mit.edu/software-datasets/FaceData2.html

9. Shmoilova R.A. Obshchaya teoriya statistiki (General Theory of Statistics), Moscow, Finansy i Statistika, 2002, 560 p.

Download