The Research Of Face Detection Based On Improved LBP Features And Floatboost

Face detection as a precondition for face recognition, whose efficiency and speed has a direct impact on the performance of face recognition. Therefore face detection has been paied more and more attention by researchs and a lot of face detection algorithms have been published. The method based on local features can be quickly and accurately determine whether there is a face in an image, but it is difficult to detect face’s size and specific location in an image. The difficulties are the variability of the human face model and external conditions, such as facial expression diversity, different lighting conditions and so on.For these reasons, there has got a larger development on the machine learning algorithms which based on the statistical theory. These algorithms have certain self-learning ability to build detection model based on the test set, and then examine the test set. In 2004, the Adaboost algorithm based on matrix features has been used in face detection by viola, which uses Adaboost learning to generate strong classifier, and achieve better result. The algorithm is one of the face detection algorithms on classic areas. However, the algorithm uses matrix characteristic which is the most original feature, in order to achieve higher detection rate, lots of feature have been chosen, and cascading series are too long, which is difficult to meet the need of real-time monitoring. The CS-LBP feature has a good representation of the partial facial texture feature proposed by Schmid, which will greatly enhance the speed of face detection. Floatboost can solve the problem that classifier obtained by Adaboost must not the best classification.Therefore, this paper combines CS-LBP feature and cascading Floatboost detection algorithm to detect face. In order to take account of TP, FP and detecting time, following improvements have been put forword: a) improve the CS-LBP texture feature, and propose a new texture feature which is simple and strong performance to reduce FP and accelerate detection speed. b) In order to further improve the performance of Floatboost, dual-threshold Floatboost has been used to train strong classify, which can further reduce the number of cascade stage and speed up the detection speed.Improved by these aspects, the result tested in MIT CBCL, BioID and AT & T database show that: the Floatboost face detection based on improved LBP features can not only reduce the detection time but also lower the FP rate while the TP rate is guaranteed to meet the demand.Finally, this paper applys the idea of Layered filter to gray image processing. That is, using algorithm which combines haar-like features and dual-threshold Floatboost cascade classifier to quickly select face candidate region, and then use the other algorithm which combines improved LBP features and dual-threshold Floatboost cascade classifier to accurately locate the face on face candidate region.