3D discrete cosine transfrom based face detection and recognition

Due to expectation of obtaining a high accuracy, interest in face recognition hasbeen shifted to the 3 dimensional (3D) methods. 3D face recognition can beanalyzed in five sub modules; face detection and alignment, feature extraction,feature selection, classification and decision. In this study the existing methodsfor each module are examined, some of them are improved, and new methodsare proposed. The whole system performance of the 3D face recognition isincreased by finding best solutions to each of sub-modules, except classification.A new Discrete Cosine Transform (DCT) based symmetry measure is presentedand used for face detection and face alignment. A 3D DCT based method forfeature extraction is proposed. For this purpose, face is represented in voxelstructure and 3D DCT is applied in two different ways; globally and locally. Inthe global approach, features are extracted from whole face or from a region ofinterest. In the local approach, using 3D sub regions, spatial information ispreserved beside frequency information. Thus negative effects of the expressionvariations are reduced. Using overlapping sub region approach, systemperformance is improved against to facial landmark shifting, such as nose andeye. Filter and sequential methods are compared to extract discriminativefeatures from 2D and 3D DCT coefficients. Variable number of coefficients isselected from each sub region and thus recognition performance is increasedwith respect to constant number of selected features from each sub region. 3Dshape and texture information are fused at data level using texture informationfor voxel values and features are extracted from this voxel structure by using 3DDCT. Existing methods and proposed methods have been successfully tested in3D RMA and FRGC databases, which are most popular databases among 3Dface recognition community. In the experimental results, high recognition ratesabove 99% are achieved by using proposed methods in related databases.