Gentle Adaboost Algorithm-based Face Detection Research

With the development of information technology, the research on face detection has been much attention in the fields of computer vision. Because the detail of face is complex and changeable, in the early period, face detection is limit to the strong constrained condition, and the detection performance can't achieve satisfactory results in the complex background. Until 2001, Paul Viola constructed a face detection system with the Adaboost algorithm, and the system has been made good results both in accuracy and speed. As a new technology of face detection, it has received widely attention.This thesis designs a face detection system with the Gentle Adaboost algorithm. The precision has been improved by modifying the training algorithm about the weak classifier. The major research work includes the following aspects:1. The thesis analyzes the whole process about face detection system including the training and detection, and introduces the Haar-rectangle feature, integral image, and the construction and training process of the classifier in detail. Finally, a simple face detection system has been constructed.2. The different rectangle features have been compared, and the number of the features has been reduced. The research results show that after reducing the number of the rectangle features, the training time has been decreased nearly two-fifths, but the performance of the system will decline.3. Two detection methods have been compared, namely: the traditional method with the pyramid and the method based on enlarging the detection windows. Regardless of the detection rate or the false accepted rate, the performance of the traditional method is better than the second method, but the detection speed of the traditional method is much slower.4. In the actual application, the cost of missing detection face is bigger than the false accepted face. According to this problem, both cost factors are introduced to the training weak classifier process, and the Cost-Gentle Adaboost algorithm has been proposed. In the training weak classifier process, this modified algorithm will select the weak classifier which owns the more classification ability to the face sample. The experimental results show that under the same conditions, the face detection system with the improved algorithm has more detection rate, but the false accepted rate is higher than the original algorithm.