Adaboost Face Detection Based on Improved Covariance Feature

Excessive number of Haar-like features and the complex threshold calculation of covariance matrix feature are two key issues of Adaboost face detection. In this paper, an efficient feature named covariance feature is proposed. The novel method divides the face image into several regions and it calculate covariance feature of any two regions. Then optimal weak classifiers will be picked out by Adaboost algorithm and they will be composed to a strong classifier. The experiments result in MIT+CMU data sets shows that the feature extraction times of the novel method is slightly slower than covariance matrix feature. However, the feature threshold is obtained much faster than covariance matrix feature, leading the significant reduction of the training time of Adaboost algorithm. Comparing with the Haar-like feature, the detection rate and speed improved obviously.