Research And Implementation Of Fast Training Algorithm Based On AdaBoost Face Detection

Face detection has important application value in information security, video monitoring, automatic face recognition, man-machine interactive and other fields. Face detection is to determine whether an image containing face by computer or other device, through using certain method to search the image. If the image contains faces, output the size and location of the faces. Face detection as an important research topic in computer vision and pattern recognition fields has been widespread concerned, because of its difficulty and no a method can achieve perfect testing results. The AdaBoost face detection algorithm proposed by Viola is a better algorithm that can attain better detecting result in the condition of real-time detection.This paper studies the AdaBoost algorithm, and proposes some improvements, finally realizes the face detection system programming. The major contribution in: firstly, In order to resolve the problem of the AdaBoost algorithm is time-consuming in training process, a new method is proposed to calculate feature's capacity of classification, so that we can improve the training speed by removing the features that have weak effect of classification. Through saving the eigenvalue-matrix that has been sorted, the training speed can be further improved. Secondly, In order to prevent the phenomenon of over-training, considering the uncertainty of sample's eigenvalues that near the threshold, we adopt a new weight eigenvalue method for these samples. Finally, in the phase of detecting faces, we firstly use cascade classifier for testing windows of faces and record them, and then confirm these windows once again through the binary chart that has been segmented by the information of complexion. As a result, we can effectively improve the detection rate.Finally, we realize a face detection system through Matlab programming. The last experimental results show that this method we use can effectively reduce the training time and achieve better detection effect.