In this Master Thesis one of the most common problems related to face detection is presented: fast and accurate unconstrained face detection. To deal with this problem a new general learning method is presented. The proposed method introduces a set of upgrades and modifications on key concepts and ideas of Decision Trees, AdaBoost and Soft Cascade learning techniques. Firstly, a new variation of Decision Trees with quadratic thresholds able to maximize the margin distance between classes is introduced. Considering a training set independent of face orientation and viewpoints information, the proposed algorithm is able to learn a combination of features to cluster faces under unconstrained face position and orientation. Next, a new definition of the Soft Cascade thresholds training principles is provided. Hence, this modification leads to a better formulation of the loss function associated to the AdaBoost algorithm. The trained face detector has been tested over the Face Detection Data Set and Benchmark (FDDB) and compared against the current state of the art classifiers. The obtained results show that the proposed face detector (i) is able to detect faces with unconstrained position, and (ii) it works faster than the current state of the art methods