Real-time face detection using a cascade of neural network ensembles
Fei Zuo · TU/e Research Portal · 2005
We propose a (neat) real-time face del.ect.or using a cascade of parallel neuTal net.work: (NN) ensembles for enhanced detection aCCUTacy and efficiency. First, we fon7/. a coordinated NN ensemble by sequentially tmining a set of neuml netw01ks with the same topology. The I.mining implicitly paTtitions the face space into a nurnbeT of di.~joint n:gions, and each NN 'is specialized in a spec~fic S11,b-l·egion. Second, to reduce the /.otal comp'lttal.ion cost for t.he face det.ect.ion, a series of NN ensembles aTe cascaded s1tch thai. the complexity of base nel:w01'ks increases. Our pTOJlosed apJlTOach achieves 1/,P to 94 % detection Tate (CM U + MIT test set) and 3-4 % frames/sec. det.ect.ion speed on a nan7/.al PC (P-IV, 3.0GHz).