ROBUST MULTIMODAL FACE RECOGNITION WITH PRE-PROCESSED KINECT RGB-D IMAGES

Abstract

Researchers have tried to improve the accuracy of face recognition by combining 2D and 3D images to overcome the problems of illumination, pose variation and occlusion. Although combining 2D with 3D have shown better results when compared with 2D images only, however applicability of these methods is inadequate in practical implementations due to high cost of 3D sensors, therefore we are using the low cost sensor Kinect acquired images. We do face recognition from RGB images, depth images and then we combine both RGB and depth maps i.e. concatenate different Modalities to improve the accuracy of recognition. Depth maps have holes and noise induced from camera sensors, therefore we process them to remove these distortions and then we apply the face recognition algorithm. Experimental results reveal that the accuracy of face recognition can be increased by combining RGB and depth images and applying pre-processing on depth maps which mitigate the effects of covariates such as holes and noise in the depth maps.