3D ear based human recognition using gauss map clustering

Abstract

This paper addresses the problem of human recognition using 3D ear biometrics. Existing feature extraction and description techniques in the literature for 3D shape recognition works well with the different class of shapes, however, not for profoundly comparable objects like human 3D ears. This work proposes an effective method utilizing Gauss mapping for feature keypoints detection and shape context to describe the detected keypoints. The proposed technique is as follows. A triangle for every point p is computed using two other points of the k-nearest neighbors within a sphere of radius r. A normal is computed for the obtained triangle and is mapped to a unit sphere. This mapping of normals is done for every conceivable triangle of point p. It is observed that mapped normals form a different number of clusters depending upon the type of surface point p belongs to. A point is considered as a keypoint if its projected normals form more than two clusters. Further, we project all the detected keypoints onto a plane and use them in the computation of feature descriptor vectors. Descriptor vector of a keypoint is computed by keeping it at the center and defining its shape context considering all other keypoints as its neighbors. To match a probe ear image with a gallery image for recognition, we compute correspondence for all the feature keypoints of the probe image to the feature keypoints of the gallery image. Final matching is performed by aligning the gallery image with the probe image and considering the registration error as the matching score. The experimental analysis conducted on University of Notre Dame (UND)-Collection J2 has achieved a verification accuracy of 98.20% with an equal error rate (EER) of 1.84%. © 2017 Association for Computing Machinery.