Download or read book Low-level Fusion and Deformation Modeling for Textured 3D Face Biometrics written by Faisal Radhi M. Al-Osaimi and published by . This book was released on 2009 with total page 152 pages. Available in PDF, EPUB and Kindle. Book excerpt: [Truncated abstract] Automatic face recognition has many crucial applications in a range of domains including human-machine interaction and security. The non-intrusive nature of face recognition is a key reason behind its suitability to such a broad range of applications. However, often these applications require high recognition accuracies and robustness which are challenging to achieve due to variations. Variations in pose, facial expression and illumination conditions may obscure the essential visual and/or geometric cues for recognition. The ability of a face recognition system to represent and match faces such that the utilized descriptive cues outweigh the associated variations is a key factor in enhancing its accuracy and robustness. This thesis presents novel algorithms that strive to achieve this goal by combining the descriptive cues in textured 3D facial scans at lower levels (data and feature fusion levels) and/or modeling the facial expression and illumination variations. An approach for combining fields of weak local and global geometrical cues (computed for each scan vertex) at the feature level was devised and used for compact representation and matching of 3D faces. The extracted fields are independent of the coordinate system in which the scan vertices were defined and hence the representation is tolerant to minor pose variations. The combined representation has demonstrated considerably high descriptiveness. In another presented approach, only local regions around key-points on the facial scans were considered. Heuristic measures for the descriptiveness of 3D local regions were defined and used for robust matching. The matching enforces consistent 3D rigid transformations and global structures among matched scans but on the other hand it allows for few mismatches among the local regions. Hence, it avoids the effects of deformed regions on matching. While the above strategies of combining information with and without the avoidance of the deformed regions gave a high recognition performance for scans under neutral expressions (minor deformations), the reported variation modeling approaches have shown higher recognition accuracies for faces with neutral or nonneutral expressions. The thesis also demonstrates that low level multimodal fusion in conjunction with variation modeling gives a higher performance than unimodal recognition and higher level fusion. Another contribution of this thesis is a novel approach for learning and morphing 3D expression deformations of 3D facial scans to those of target scans. This neutralizes the effects of expression variations among matched scans. In an extension of this approach, unseen facial scans under any facial expression were decomposed into estimates of neutral 3D faces and deformation residues. This is a very challenging problem due to the lack of the target scan. The estimated neutral faces and the deformations residues were used for face recognition and expression classification, respectively. The thesis also presents a novel approach for illumination normalization of the facial texture in textured scans...