Face Location And Facial Feature Extraction In Complex Background

Finding huamn face automatically in a scene is a difficult but significant problem. In my Ph.D thesis, we will study it in 4 aspects. 1. Propose an improved hierarchical knowledge-based method to locate human faces in a complex background. It consists of three levels up to down. (1) At the highest level, we present a new method based on quartet (4×4) and 5×3 binary mosaic images at different resolutions for locating face candidates. (2) At the second level the candidate region is analyzed by the Octet (8×8) mosaic image, the decision will be made whether the candidate is a face or not. (3) At the lowest level, a novel approach is proposed and the parts of a face are verified by a high-boost filter, projection analysis and the eye corner detection. 2. Introduce an improved down-top method for detecting the face and facial features. (1) First we use a new eye pair detector to identify the possible eye pair as the reference parts of a face by processing the 5×5 binary mosaic images at multi-resolution and marking symmetry blocks as eye pairs. (2) Second A face candidate region with eye pair marked at the first step is matched by a whole face detector, which is referred as recognition stage. Based on biologically motivated image, face region is represented by a radially non-linear sampling lattice. Each sampling node is surrounded by a set of adjacent nodes which receive input from the same patch of the visual field. The outputs of the nodes are fed to SOFM modules. Optimal set of face are found out by an enhanced SOFM approach. 3. Propose a new filter to find out frame(structual) curves in an image by a feature inertia surface method to locate boundaries of objects approximately, and verify if it is a face. A novel approach is also suggested to decide lines reserved for frame curves. It is also a three level approach. (1) At the first level, calculate feature inertia surface by a filter bank referred as differences-of-Gaussian(DOG) and differences-of-offset-Gaussian(DOOG). A new filter is also proposed to join the filter bank and improve the scheme. The image passes through this filter bank, followed by a non-linear inhibition, localized in space within and among the filter response profiles. Thresholding is to suppress weak response near the stronger one. The feature inertia surface is obtained after smoothing operation and gradients composition. (2) At the second level, locate and extract frame curves using the feature inertia surface as input. It is done at following two steps: Step A: Define 8-direction reference lines, and perform the next two operations to each direction: assign parallel lines with the same slope as the reference line of the direction, the densities of these processing lines at each point are determined by the feature inertia surface; Combine and merge the parallel lines according to the distances between lines. Step B: Combine 8-direction processing lines to form a set of closed curves referred as frame curves. (3) At the third level, perform the facial feature verification. 4. We also propose another face segmentation method using the concept of synergetics and a face deformable template. According to the concept of synergetics, pattern formation and pattern recognition are formallyanalog. In a pattern recognition process, one can consider a deformation between the stored prototyped pattern and the test pattern to be recognized. We use a deformable template to represent a face, and derive a gradient dynamics function as a transformation function, a cost is calculated to check the deformable ability of a propotyped pattern to an input test pattern and judge if the input pattern is a face. After segmentation of a face is finished, initial position of facial features are decided by integrated projections of a binary boost_filtering image, and the feature parameters of their deformable template are accurately estimated through energy functions'optimization. Features are also verified by isodensity maps. Clustering based Image restoration and rules of virtual driving system are also discussed in this paper.