The Research Of Frontal Face Detection Based On Support Vector Machine

Recently, the research of Face Detection focuses on Face Detection under complex background. One aspect is cutting down false alarms and raising face detect rate. Another aspect is speeup while easy to train. Since SVM has solid mathematic background and good generalize ability, it has been widely used in Pattern Recognition and Regression Analysis. In this paper, we mainly discuss the application of SVM in Face Detection.SVM is a Statistical Learning Theory, we explained its theory through a simple linear classification problem. Then we introduced a few kinds of improved algorithms, which can greatly increase the training speed of SVM.Traditional SVM Face Detection methods and their Coarse Filter take pixels as input. On the basis of experiment, we compared three kinds of Coarse Filters - Template Matching, PCA, Sampling Non-Linear SVM, the last one achieved the best performance. When detecting, all windows in scaled image series pass through the coarse filter and the main classifier, then merge, locate and output to original image.We proposed a new Face Detection method that combined Rectangle Feature, Cascade Classifier and SVM. We took a simplified Cascade Classifier as a Coarse Filter, and use Rectangle Feature as the input of SVM main classifier. Compare with Viola's method, although the detecting speed is relatively slower, the training speed had been greatly increased, while achieved the same face detect rate and false positive rate.In order to further increase detecting speed, we introduced the application of SIMD (single-instruction, multiple-data) in Face Detection. By using different instruction set, we can process four integer or single float operands in one instruction.In the final part, we discussed the possible ways of developing Face Tracking system based on SVM, and the possible approaches to speedup cascade classifier training by using feature selection.