Face Recognition Algorithms Based On DCT

Face recognition is an important technology for biometrics based recognition, which is widely used in airport and other key occations as personal identification, validation and safety surveillance,etc. Face recognition includes three parts: preprocessing, feature extraction and classification. This thesis presents the studies of preprocessing and feature extraction in face recognition.Discrete cosine transform(DCT) is an excellent method for data compression. In the proprecessing part, we first introduce the application of DCT to reduce dimension: limited coefficients at the up-left-corner of DCT matrix contain most information of an image which are more useful for recognition, then dimension reduction can be achieved by only keeping this part of coefficients. Then we approse a preprocessing algorithm for illumination compensation based on DCT, which takes advantage of characteristics of DCT: certain coefficients work for half light, so we can compensate the illumination by dealing with these coefficients.Fisher linear discriminant analysis has found successful applications in face recognition. We approve two DCT and LDA based algorithms: D-LDA method and Null space method based on DCT. In order to save time spent on recognition, we prove that D-LDA and null space method can be used on DCT domain directly. For D-LDA ,we use DCT to reduce dimension before doing eigen analysis on between-class scatter matrix to collect information, and the information in null space of within-class scatter matrix is saved. For Null space method, we use DCT instead of "pixel grouping" which is too complex and much useful information is lost because of its application. For both methods, we propose a strategy to compute the total between-class scatter matrix and the pooled within-class scatter matrix using some weight functions: A weighted total between-class scatter matrix is constructed in which smaller distances are more heavily weighted than larger distances, because those classes which are clustered together are more likely to be misclassified; a weighted pooled within-class scatter matrix is constructed in which outlier class is more lightly weighed, because if one class is well separated from the other classes in the input space ,then whether the within-class covariance matrix of this class in the new space is compact or not will not have much influence on the classification result. Experiment results show that this strategy has a better performance.In order to improve the recognition result, we introduce an iterative linear discriminant analysis-fractional linear discriminant analysis to face recognition. F-LDA can not be used in recognition because of the complicated calculation. In this thesis, the dimension is small enough in the step of feature extraction, so F-LDA can be used without any trouble.The algorithms proposed in this thesis are experimented on ORL and Yale databases. The experiment result show that the proposed methods can not only increase the recognition rate and save much space for storage, but also reduce recognition time prominently.