Iannarelli’s studies demonstrated that ear shape represents a biometric identi?er able to authenticate people in the same way as more established biometrics, like face or voice for instance. However, not many researches can be
found in literature about ear recognition. In most cases existing algorithms are borrowed from other biometric contexts. An example is PCA (Principal Component Analysis). Eigen-ears only provide high recognition rate in closely controlled conditions, while performances decay even for small changes in environmental conditions. The ear biometric provides best performances, because it does not undergo to meaningful variations (i.e. expressions, occlusions,...). There are very few works on the ear biometric in literature. Indeed, only two significant techniques have been investigated:
- Iannarelli’s method: a limited number of control points is located on the ear and a set of distances among these points is built. These distances make up the feature vector.
- force fields: each pixel into the ear image is considered as an electric charge, wielding influence (electric force) over all the others. Force fields are derived summing all these forces, which are represented by vectors. All the force fields converge in some points called wells. Just next to each well a several statistic measures are drawn out, giving the feature vector.
We propose a fractal based technique, namely HERO (Human Ear Recognition against Occlusions) to classify human ears. The feature extraction process has been made local, so that the system gets robust with respect to small changes in pose/illumination and partial occlusions.
On the contrary, at BIPLab we readjusted a new efficient and original technique based on the fractal theory to the case of the human ears. The ear is segmented in four interest regions of equal size (upper-left, upper-right, lower-left and lower-right corners): each of such regions is processed independently so to make the algorithm robust to different variations, including occlusion.
Features extracted from each region are then chained in a single vector representing the biometric key for the ear. Experimental tests have been conducted in order to assess the performance of HERO with respect to small variations in pose/illumination and presence/absence of occlusions. Four methods already used in face recognition have been selected to compare them and assess HERO’s performances: 1) the Principle Component Analysis (PCA); 2)the Linear Discriminant Analysis (LDA); 3) the Kernel machine-based Discriminant Analysis (KDA) method; 4) the Orthogonal Locality Preserving Projections (OLPP). Two sets of 256 × 256 pixels ear images have been extracted from the two largely used databases Notre Dame and FERET. The Notre Dame database consists of pro?le images of 114 subjects. The subset we used consists of two pictures for each subject (228 ear images), the former acquired in controlled conditions (gallery image) and the latter under different pose/illumination conditions. The FERET ear subset has been obtained by selecting 200 pro?le faces of the ?rst 100 subjects (with pro?le images). Experimental results show the robustn