Expression and Occlusion invariant 3D face recognition based on region classifier

In recent year, 3D face images play a major role in face recognition over corresponding 2D images. In this work, the 3D range images are used for face recognition based on region classifiers. Three different regions: eye, nose, and mouth separately classify a face image locally. At first, local binary patterns (LBP) are calculated for all pixels on face images. A new image formed with these LBP values are cropped and then divided into three horizontal regions, namely eye, nose, and mouth. For each such region histogram of oriented gradient (HOG) is used for feature vector creation. Two databases: Frav3D of 106 different subjects and our database consisting of 102 various individuals are used for recognition. Only frontal images with occlusion, expression and neutral images from those databases are utilized for this proposed system. Two fold cross validation technique with a nearest centroid-based classifier is used for classification. In the case of decision level fusion, recognition accuracy is 88.86% on Frav3D database and 77.5% for our newly created database. On the other hand, score level fusion has shown 78.5% recognition accuracy for FRAV3D database and 65.55% for our database.