Research Of Maximum Margin Criterion Based On Cost Sensitive

Traditional subspace learning based face recognition methods try to seek a low dimensionalsubspace for classification In the process of face recognition. These methods usually assume thatthe misclassification cost between different classes is the same. However, this assumption is notreasonable in some certain scenarios, since that different misclassification results may contribute tovarious cost. Therefore, how to improve recognition performance by considering themisclassification cost issue is an important topic, which also is the focus of this thesis. Recently,some cost-sensitive subspace learning methods have been addressed, including CSPCA, CSLPP andCSLDA. But these methods always suffer from some certain problems. Although CSPCA andCSLPP are two representative methods, their discriminative capability limits because they don’tutilize the discriminant information. And CSLDA, a supervised method, also may encounter thesmall-size-sample problem when the number of samples is less than the dimension of samples.Maximum margin criterion (MMC) is an effective supervised subspace learning method. SinceMMC does not contain the matrix inversion operation for within-class matrix, it avoids thesmall-size-sample problem. By considering the misclassification cost issue, this thesis first proposesthe cost-sensitive maximum margin criterion (CSMMC). We incorporate the cost-sensitive issueinto both within-class matrix and between-class matrix, and design the cost-sensitive basedwithin-class matrix and between-class matrix. By using the eigenvalue decomposition process, wecan obtain the projection transformation. It should be noted that the parameters of cost-sensitiveissue are determined according to different classes by choosing the values that can yield the bestrecognition performance.CSMMC needs to reshape the original sample into one vector, which may destroy the originallocal structure of original sample like face image. Moreover, CSMMC yields high dimensions forthe samples, and then brings the singularity problem for scatter matrixes. To solve this problem, thisthesis further proposes2-dimension cost-sensitive maximum margin criterion (2DCSMMC). Byintroducing the cost-sensitive technique, this approach employs the original samples to construct thescatter matrixes without the operation of reshaping them into one vector.Traditional linear feature extraction methods reduce the dimension of pattern samples throughlinear transformation. But when original samples are non-linear distribution, the existing linearmethods are difficult to extract effective identification features. Then we can convert it to a new,more high-dimensional space makes original samples linearly separable. Based on the theory ofkernal, we propose the kernel cost-sensitive maximum margin criterion (KCSMMC). We firstproject the low dimensional samples into high dimensional space through the kernel function. Thenwe construct the scatter matrixes by incorporating the misclassification cost issue. And the solution can be obtained with the kernel theory.Extensive experiments on three widely-used face databases, including AR, FERET andCAS-PEAL, verify the effectiveness of our proposed approaches.