Face detection based on multi-scale enhanced local texture feature sets

This paper presents a distinctive rectangle feature Multi-Scale Local Ternary Patterns (MS-LTP) for face detection. The MS-LTP is a generalization of the Local Ternary Patterns (LTP) [1] and is able to capture larger scale structures of faces. It´s less sensitive to noise and more discriminative that can reduce the number of weak classifiers for the AdaBoost learning algorithm to construct a strong face/non-face classifier. The size of the MS-LTP feature set is also medium for the AdaBoost learning algorithm to select a proper set of features. Our experimental results on the CMU-MIT frontal face test set show that the MS-LTP outperforms Haar, Local Binary Patterns (LBP) under noisy conditions and the MS-LTP based face detector works more rapidly.