Heterogeneous Parallel Ensemble Classifiers for Face Detection

1352 Words3 Pages

Heterogeneous Parallel Ensemble Classifiers for Face Detection Abstract - Face detection is a process to determine the presence of a human face in a given image, and if exists then returns the location of detected face. Face detection is a difficult process due to certain factors, such as background complexity, illumination conditions, scale, expression, position, orientation and pose of the face. The presence of external facial features such as glasses, moustaches and beards further complicates the process of face detection. Face detection acts as a very first and an important step in applications such as face recognition, human computer interface, automatic surveillance and content-based image retrieval. In this paper, we proposed a face detection technique based on ensemble-classifiers, which employs multiple heterogeneous classifiers in parallel fashion. The classifiers may itself be ensembled in cascaded manner. The seminal work of Adaboost based face detection has shown robust classification capability while some of the non-face patterns are identified as face patterns. We will mainly focus on reducing the false positives produced by the Adaboost based face detection system by cascading Adaboost with (i) Neural Networks, (ii) Support Vector Machines and (iii) Bayesian Networks. Then all these three-cascaded configurations will be ensembled in a parallel fashion. We will be using eager learners (ANN, SVM, Bayesian networks etc) as they take more time for training but less time in prediction. Hence once trained eager learners predict quickly and are expected not to effect negatively on time efficiency of Adaboost algorithm. I. Introduction: Face identification is a two fold problem which involves face detection an... ... middle of paper ... ...y the weak classifier are assigned lower weights, whereas the examples that are misclassified are given higher weights in a moving probability function which defines the training samples. The random training becomes more deterministic, focusing on the examples that are difficult to classify. Adaboost algorithm for the two class problem is explained as follows. Given (xi; yi), … , (xm, ym) xi Є X; yi Є Y = {-1; +1} where X is the training set and Y is the label set. Weight distribution over X D1(i)=1/m Є (1) where i = 1; :::;m and D(i) is the set of weights over the training set. Initially all weights are set equally.The weight of this distribution on training example i on round t is denoted by Dt(i). _ For t = 1; :::; T, – Train a weak classifier h using a learning function L on samples X and weight distribution D ht(x) = L(X;Dt

More about Heterogeneous Parallel Ensemble Classifiers for Face Detection

Open Document