The Face Detection Of AdaBoost Algorithm Based On Feature Fusion

Face detection is a key step in face recognition. It is widely applied in computer vision and pattern recognition. The human face is a characteristic non-rigid structure,it is susceptible to external environmental factors. So face detection is a complex and challenging task. But the applications of face detection are very widespread, so all kinds of face detection algorithms are proposed. Ada Boost algorithm has higher detection rate and detection speed, making it a hot topic in many domestic and international researchers. Therefore how to improve the detection rate and reduce the training time becomes the main research goal. This thesis is based on Haar-like feature and the traditional Ada Boost algorithm, improve the method of feature extraction and the steps of Ada Boost algorithm. These improvements are simulated on the platform of Matlab, confirm the improved Ada Boost algorithm proposed has obvious improvement.The main works of this thesis are listed as follows:Because of the shortcomings of Haar-like feature, introduces another feature(LBP features), and proposes an improved method of Haar-like features and LBP feature fusion, this method is introduced into the Ada Boost algorithm in order to achieve higher detection rate. Firstly, LBP feature is made some improvements to describe the human face better. Secondly, this thesis extracts Haar-like features and Local Binary Pattern features in certain proportion of training samples. All the features are applied to train the weak classifiers. Choose the best weak classifiers which make the minimum classification error, and integrate these best weak classifiers linearly to minimize the two objective functions. The experiment result showed that although the detection rate of this method is similar to the traditional Adaboost algorithm, the method in this thesis can effectively shorten the training time, and lessen the number of features.Some rare samples such as noise samples are included in the training samples,they usually result in the overfitting of sample weights. So an improved Ada Boost algorithm based on changing the weight parameters is proposed. This method improves weight updating rule and weights normalization rule of the traditional Ada Boost algorithm. Face samples and non-face samples are normalized respectively, and propose a weight update rule associated with false detection rate.The method uses Haar-like features and LBP feature fusion instead Haar-like feature for feature extraction. Then the improved weight normalization rule and weights update rule are introduced into Ada Boost algorithm. The experiment result shows that compared to Ada Boost algorithm based on Haar-like features, the improved Ada Boost algorithm can effectively reduce the overfitting of sample weights and improve the detection rate. And the performance of face detection improves obviously.