Research On The Algorithms Of Fast Face Detection And Feature Extraction

Face detection and feature extraction are two important research branches of patternrecognition region. The correlative theory and technology with. face detection and featureextraction is not only the basis of face Recognition, but also the key to resolve manyquestions of object detection and pattern recognition, such as vehicle detection, passerbydetection, character recognition and others biometric recognition et al. In addition, theresearch of face detection and feature extraction has an effect on the development ofacademic subjects including pattern recognition and machine vision. Therefore, theresearch of face detection and feature extraction has very important business and sciencevalue. On the one hand, the paper carried through some research to the current leading fastface detection methods and some face detection algorithms are proposed to enhance theperformance of face detection. On the other hand, many methods that extracting efficientface features are developed after the paper investigated main methods of feature extraction.AdaBoost and Cascade algorithm are two poplar methods of training face detector.Because mass Haar-Like features are using in the training process, it will consume a lot oftime and memory. The paper presented a kind of Walsh feature which has localorthogonality and replace Haar-Like features with these Walsh features in the trainingprocess, which can decrease the redundancy among the features. The experimental resultsshow that less Walsh features can obtain good face detection performance. Aimed at theshortage of entirely inheriting prior classifiers in the Nesting Cascade algorithm, anenhanced cascade algorithm with independence characteristic and inheritance specialty isproposed. The experimental results on MIT-CBCL database show that Enhanced Cascadealgorithm can increase the test precision. In addition, the paper analyzed the regiondivision problem of the Real AdaBoost and developed the extended Real AdaBoost basedon the weight histogram with the fuzzy region. The experimental results on MIT-CBCLdatabase, demonstrated the algorithm's validity. Finally, several detectors trained by theabove proposed algorithms are used to detect face in the MIT+CMU frontal face test setand CMU profile face test set and the detected results demonstrate that they are moreeffective than the other detectors.The Maximal Rejection Classifier (MRC) is a new classification method on the basis ofFisher Discriminant Analysis. MRC-Boosting that integrating Boosting and MRC enhancesthe classifier's performance much more. Aimed at some disadvantages among theMRC-Boosting and MRC, an Adaptive Maximal Rejection Discriminant Analysis(AdaMRDA) is proposed. The method that obtained orthogonal projection vectors ofAdaMRDA is given and an adaptive weight adjustment and classification menthod whichenhances the classifying performance is designed in the paper Some experimental resultson MIT-CBCL and FERET database show that AdaMRDA is more effective thanMRC-Boosting and MRC. In addition, according to the respective merits of Real AdaBoostand AdaMRDA algorithm, a fast face detection based on the AdaBoost+AdaMRDAalgorithm is proposed and the algorithm obtained good face detection result in the MIT+CMU test set.Principal Component Analysis (PCA) is the well-known method in pattern recognition,but extracted features aren't very efficient to classification based on classical PrincipalComponent Analysis. Based on the image retrieve principle, the paper presents a kind ofRetrieve Space Principal Component Analysis (RS-PCA). Then, Supervised RetrieveSpace Principal Component Analysis (SRS-PCA) using classificatory information isdeveloped according to RS-PCA. The algorithm makes the extracted features moreeffective and the recognition precision is increased. The experiments resulted on ORL andYale face database demonstrate that the proposed algorithms has more powerful andexcellent performance than classical principal component analysis. Then, combining theideas, of 2DDPCA, some Supervised Retrieve Space 2DDPCA algorithms are developedand the experimental results demonstrate their validity. Finally, a kind of PCA based on themulti-scale SVD feature is proposed. The experimental results show that the algorithm canimprove face recognition performance.Linear Discriminate Analysis (LDA) is an effective method in pattern recognition too.A Multi-band Linear Discriminate Analysis is proposed, by which LDA is established onthe whole samples space. Based on MBLDA, the data loss using PCA+LDA is avoided andthe recognition performance is improved. An Intensified Linear Discriminate Analysis(ILDA) criterion, which integrates PCA and LDA merits sufficiently, is proposed in thepaper. Many experiments on ORL, Yale and NUST603 indicate that ILDA is effectivefeature extraction method. In addition, a fast DLDA algorithm is developed to remedy theshortage of classical DLDA that consumed a lot of time and space in processing the highdimension samples. On the basis of the fast DLDA, a DLDA algorithm based on themulti-scale low-frequency feature is proposed. The experimental results show that thealgorithm can extract effective face features which can obtain higher face recognitionperformance.Finally, a new method, called Expected Discriminant Analysis (EDA), is proposedbased on the similarity of samples' distribution after some disadvantages of LDA areindicated. What's more, the method of resolving the optimal Discriminant vectors with theorthogonal and approximate uncorrelated specility is presented. Some experiments on twodifferent databases demonstrate that the proposed method is better than the other methods.In addition, combined the kernel method, a Kernel Expected Discriminant Analysis(KEDA) is developed and a fast kernel projection method is proposed with same trainedresult. Some experimental results on ORL face database show that the proposed algorithmis effective.