Abstract
dc:descriptionIn this thesis, an automatic face recognition system is developed which is invariant against varying facial expressions and poses. This is an essential condition for the application as a non-intrusive person recognition system.Existing face recognition systems require a frontal, neutral view of faces (norm-face) to perform a robust classification. To achieve a robust recognition using these methods under varying head orientations and facial expressions, a sophisticated pre-processing of general facial images is introduced which is capable of transforming a general facial image into a norm-face image using a model-based normalization approach. For generating the norm-face image the size and the position of the face is determined using an efficient, holistic detection algorithm. Afterwards, facial landmarks are localized using a statistical face model that combines textural information as well as the position of the facial landmarks extracted from sample (facial) views. By applying an iterative fitting algorithm, the values of the model parameters are determined so that the model represents the original facial image in an optimal way. The position of the facial landmarks in the processed image can then be derived. The subsequent, two-stage head model fitting process is based on the definition of correspondences of multiple facial features between the 2D image and the 3D model. By optimizing the parameters of an affine transformation the model is initially adapted to the head pose. The optimization process minimizes the distances between the 2D facial landmarks and the corresponding, projected 3D model vertices. In the second step, a distortion of the head model using a locally effective RBF interpolation is performed for a more precise modeling of the individual facial proportions and the shown facial expression. After the fitting process, the model is textured, re-transformed into its original reference form and the image information of a possibly occluded face half is reconstructed. The final projection into a suitable image plane results in a neutral frontal view. The performance of this normalization approach was comprehensively evaluated. From all detailed results two main conclusion can be derived: A recognition performance of 89,4% is achieved when the pose deviation is up to a maximum of 45° horizontally (left-right) and 20° vertically (up-down) when different facial expression are present at the same time. This was measured on a database of 12 persons using a single training image for each person and in total 1824 test images (surveillance scenario). This states a significant performance increase in comparison to the recognition without face normalization (42,2%). Are solely facial expression changes present in the test images and neutral views are used for training, the recognition rate is 98,4% (passport scenario, 21 persons and three training images per person). The presented normalization approach using a single, generic 3D head model is an efficient and robust method for pose and facial expression invariant face recognition which in comparison to existing approaches does not require any manual intervention. Especially the automatic and combined processing of facial expressions and head poses in the considered range without usage or generation of vast image databases is innovative and provided by the presented approach for the first time.
Degree
thesis:*- Grantor dc:publisher
- Hut
- Year dc:date
- 2008
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hähnel, Michael
- Contributors dc:contributor
-
- Kraiss, Karl-Friedrich
Subjects
dc:subject × 14Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- ger