A robust analysis, detection and recognition of facial features in 2.5D images

Abstract

A robust technique for recognition of 3D faces which performs well with face images with various poses, expressions and occlusions. In this method, the face images represented in 3D mesh format are smoothed using trilinear interpolation and then converted to 2.5D image or range images. Nose-tip which is the most prominent feature on human face is detected first on the corner points selected by 3D Harris corner and curvedness at those corner points. K-Means clustering is applied to group those corner points in 2 groups. The cluster of points with larger curvedness values represents the possible locations of nose-tip. Nose-tip is finally localized using Mean-Gaussian curvature values of the prospective corner points in that cluster. Using the nose-tip location, other facial landmarks namely corners of the eyes and mouth are located and a facial graph is generated. The dimensionality of 2.5D feature space is that, depth values are stored at each (x, y) grid of the 2.5D image, so a 3D face image uses some function to map the depth value at any pixel position to the intensity with which that pixel will be displayed. Here finally extracted features for each subject is of dimensionality [1ź?ź21], taking into account the Euclidean distances in three dimensional form between each feature points detected automatically. Taking Euclidean distances between all pairs of landmark points as features, face images are classified using Multilayer Perceptron (MLP), as well as Support Vector Machines (SVM). Maximum recognition rates of 75 and 87.5 % have been obtained in case of Bosphorus Databases, 62.5 and 87.5 % in case of GavabDB databases, 75 and 87.5 % in case of Frav3D Databases by Multilayer Perceptron and Support Vector Machines respectively.