Anika Pflug

Ear biometrics: a survey of detection, feature extraction and recognition methods

The possibility of identifying people by the shape of their outer ear was first discovered by the French criminologist Bertillon, and refined by the American police officer Iannarelli, who proposed a first ear recognition system based on only seven features. The detailed structure of the ear is not only unique, but also permanent, as the appearance of the ear does not change over the course of a human life. Additionally, the acquisition of ear images does not necessarily require a persons cooperation but is nevertheless considered to be non-intrusive by most people. Owing to these qualities, the interest in ear recognition systems has grown significantly in recent years. In this survey, the authors categorise and summarise approaches to ear detection and recognition in 2D and 3D images. Then, they provide an outlook over possible future research in the field of ear recognition, in the context of smart surveillance and forensic image analysis, which they consider to be the most important application of ear recognition characteristic in the near future.

A comparative study on texture and surface descriptors for ear biometrics

Recent research in texture-based ear recognition also indicates that ear detection and texture-based ear recognition are robust against signal degradation and encoding artefacts. Based on these findings, we further investigate and compare the performance of texture descriptors for ear recognition and seek to explore possibilities to complement texture descriptors with depth information. On the basis of ear images from visible light and depth maps, we extract texture and surface descriptors. We compare the recognition performance of selected methods for describing texture and surface structure, which are Local Binary Patterns, Local Phase Quantization, Histograms of Oriented Gradients, Binarized Statistical Image Features, Shape Context and Curvedness. Secondly we propose a novel histogram-based descriptor that performs feature level fusion by combining two information channels to a new feature vector. Our concept can either be applied for fusing two different texture or two different surface descriptors or to combine texture and depth information. Based on the results of the previous experiment, we select the best performing configuration settings for texture and surface representation and use them as an input for our fused feature vectors. We report the performance of different variations of the fused descriptor and compare the behavior of the fused feature vectors with single channel from the first series of experiments.

2D ear classification based on unsupervised clustering

Ear classification refers to the process by which an input ear image is assigned to one of several pre-defined classes based on a set of features extracted from the image. In the context of large-scale ear identification, where the input probe image has to be compared against a large set of gallery images in order to locate a matching identity, classification can be used to restrict the matching process to only those images in the gallery that belong to the same class as the probe. In this work, we utilize an unsupervised clustering scheme to partition ear images into multiple classes (i.e., clusters), with each class being denoted by a prototype or a centroid. A given ear image is assigned class labels (i.e., cluster indices) that correspond to the clusters whose centroids are closest to it. We compare the classification performance of three different texture descriptors, viz. Histograms of Oriented Gradients, uniform Local Binary Patterns and Local Phase Quantization. Extensive experiments using three different ear datasets suggest that the Local Phase Quantization texture descriptor scheme along with PCA for dimensionality reduction results in a 96.89% hit rate (i.e., 3.11% pre-selection error rate) with a penetration rate of 32.08%. Further, we demonstrate that the hit rate improves to 99.01% with a penetration rate of 47.10% when a multi-cluster search strategy is employed.

Feature extraction from vein images using spatial information and chain codes

The pattern formed by subcutaneous blood vessels is unique attribute of each individual and can therefore be used as a biometric characteristic. Exploiting the specific near infrared light absorption properties of blood, the capture procedure for this biometric characteristic is convenient and allows contact-less sensors. However, image skeletons extracted from vein images are often unstable, because the raw vein images suffer from low contrast. We propose a new chain code based feature en- coding method, using spatial and orientation properties of vein patterns, which is capable of dealing with noisy and unstable image skeletons. Chain code comparison and a selection of preprocessing methods have been evaluated in a series of different experiments in single and multi-reference scenarios on two different vein image databases. The experiments showed that chain code comparison outperforms minutiae-based approaches and similarity based mix matching.

Ear Detection in 3D Profile Images Based on Surface Curvature

Although a number of different ear recognition techniques have been proposed, not much work has been done in the field of ear detection. In this work we present a new ear detection approach for 3D profile images based on surface curvature and semantic analysis of edge-patterns. The algorithm applies edge-based detection techniques, which are known from 2D approaches, to a 3D data model. As an additional result of the ear detection, the outline of the outer helix is found, which may serve as a basis for further feature extraction steps. As our method does not use a reference ear model, the detector does not need any previous training. Furthermore, the approach is robust against rotation and scale. Experiments using the 3D images from UND-J2 collection resulted in a detection rate of 95.65%.

Robust localization of ears by feature level fusion and context information

The outer ear has been established as a stable and unique biometric characteristic, especially in the field of forensic image analysis. In the last decade, increasing efforts have been made for building automated authentication systems utilizing the outer ear. One essential processing step in these systems is the detection of the ear region. Automated ear detection faces a number of challenges, such as invariant processing of both left and right ears, as well as the handling of occlusion and pose variations. We propose a new approach for the detection of ears, which uses features from texture and depth images, as well as context information. With a detection rate of 99% on profile images, our approach is highly reliable. Moreover, it is invariant to rotations and it can detect left and right ears. We also show, that our method is working under realistic conditions by providing simulation results on a more challenging dataset, which contains images of occluded ears from various poses.

Impact of severe signal degradation on ear recognition performance

We investigate ear recognition systems for severe signal degradation of ear images in order to assess the impact on biometric performance of diverse well-established feature extraction algorithms. Various intensities of signal degradation, i.e. out-of-focus blur and thermal noise, are simulated in order to construct realistic acquisition scenarios. Experimental evaluations, which are carried out on a comprehensive database comprising more than 2,000 ear images, point out the effects of severe signal degradation on ear recognition performance using appearance features.

Segmentation and Normalization of Human Ears Using Cascaded Pose Regression

Being an emerging biometric characteristic, automated ear recognition is making its way into forensic image analysis for law enforcement in the last decades. One of the most important challenges for this application is to deal with loosely constrained acquisition scenarios and large databases of reference samples. The research community has come up with a variety of feature extraction methods that are capable of handling occlusions and blur. However, these methods require the images to be geometrically normalized, which is mostly done manually at the moment.

Effects of severe signal degradation on ear detection

Ear recognition has recently gained much attention, as for surveillance scenarios identification remains feasible, in case the facial characteristic is partly or fully covered. However video footage stemming from surveillance cameras is often of low quality. In this work we investigate the impact of signal degradation, i.e. out-of-focus blur and thermal noise, on the segmentation accuracy of automated ear detection. Realistic acquisition scenarios are constructed and various intensities of signal degradation are simulated on a comprehensive dataset. In experiments different ear detection algorithms are employed, pointing out the effects of severe signal degradation on ear segmentation performance.

Towards making HCS ear detection robust against rotation

In identity retrieval from crime scene images, the outer ear (auricle) has ever since been regarded as a valuable characteristic. Because of its unique and permanent shape, the auricle also attracted the attention of researches in the field of biometrics over the last years. Since then, numerous pattern recognition techniques have been applied to ear images but similarly to face recognition, rotation and pose still pose problems to ear recognition systems. One solution for this is 3D ear imaging. the segmentation of the ear, prior to the actual feature extraction step, however, remains an unsolved problem. In 2010 Zhou at al. have proposed a solution for ear detection in 3D images, which incorporates a nave classifier using Shape Index Histogram. Histograms of Categorized Shapes (HCS) is reported to be efficient and accurate, but has difficulties with rotations. In our work, we extend the performance measures provided by Zhou et al. by evaluating the detection rate of the HCS detector under more realistic conditions. This includes performance measures with ear images under pose variations. Secondly, we propose to modify the ear detection approach by Zhou et al. towards making it invariant to rotation by using a rotation symmetric, circular detection window. Shape index histograms are extracted at different radii in order to get overlapping subsets within the circle. The detection performance of the modified HCS detector is evaluated on two different datasets, one of them containing images n various poses.