Robust Face Detection Using One-Class Estimation and Real Adaboot

Abstract

We propose a robust face detection algorithm using one-class estimation and Real AdaBoost. Inspired by the first practical face detection algorithm by Viola and Jones, many varieties of face detection algorithms have been proposed. The common feature of their algorithm is a cascaded structure of combined Haar-like features trained by a boosting algorithm. Of course this framework has achieved a successful result of high detection rate and low false positive rate in a short time and has been applied to many imaging products. But because the non-face class includes multiple sub-classes and their variations are too many to be collected and covered in training data, unexpected false positives inevitably happen in the real world data. That is a problem of self-printing systems for digital cameras because they need to handle all kinds of pictures in the real world. Furthermore because they use detected face regions for image enhancement before printing, to suppress false positives is a big issue of self-printing systems.To solve the problem of false positives in the real world, we model a non-face class using one-class estimation of faces, and developed a new face detection algorithm combining one-class estimation and a cascaded face detection by Real AdaBoost. As a result of the experiment using pictures of digital cameras, we achieved about twice faster face detection with eight times lower false positives than a conventional cascaded face detector, and also more precise face size detection.