Multiple Kernel Learning Algorithm And Its Application To Face Detection

During face recognition process, face detection is the premise and foundation of face recognition, face detection result has a direct impact on the accuracy of face recognition. The main work of face detection finds out human face from the still image or a video sequence, detects the position and marks the size of the face, face detection process can be divided into: a face image acquisition, preprocessing of face images, feature extraction and detect the presence of human faces. Which detects the presence or not can be seen as a binary classification problem, for binary classification problem, SVM method has been proposed by Vapnik and his colleagues, but SVM also has some limitations in large-scale samples, irregular data distribution or distribution in high-dimensional feature space when uneven. To compensate for the defects SVM learning, Multiple Kernel Learning(MKL) has been proposed by Lanckriet and his colleagues, the essence of MKL gets convex combination, namely, how to confirm the weights of the basic kernel functions, finally, the problem becomes that how to choose the right kernel function and their coefficients.Research works of this paper boost the computation of the efficiency and accuracy of classification based on multiple kernel learning. Using improved gradient descent algorithm to solve JHMKL and OWMKL model proposed, and the simulation experiments has been done with four UCI databases and three individuals face database, its rationality has been verified. The main works of this paper:(1) Because MKL exists high computational complexity problem, using improved gradient descent algorithm to solve original MKL model, then a description of MKL model is given based on improved gradient descent algorithm.(2) For MKL classification accuracy question, using Jaakkla-Haussler bound, Opper-Winther bound, JHMKL and OWMKL model have been proposed, and then, we use improved gradient descent algorithm to solve JHMKL and OWMKL models. OWMKL model JHMKL steps have been given, and finally we simulate in UCI datas, compared with SVM and MKL, JHMKL and OWMKL algorithm has higher classification accuracy, compared with MKL, they have less running time.(3) For feature extraction part of face detection, using a combination of kernel function instead of the original single kernel, which made multiple kernel PCA feature extraction method, and this result as a input of JHMKL and OWMKL. Using Yale, ORL, Caltech 101 image sets for face detection simulation, compared with SVM and MKL, JHMKL and OWMKL higher average detection accuracy rate, which illustrates proposed model is reasonable and effectiveness.