This PDF file contains the front matter associated with SPIE Proceedings Volume 7667, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.

This paper presents a local feature-based method for matching facial sketch images to face photographs, which is the first known feature-based method for performing such matching. Starting with a training set of sketch to photo correspondences (i.e. a set of sketch and photo images of the same subjects), we demonstrate the ability to match sketches to photos: (1) directly using SIFT feature descriptors, (2) in a "common representation" that measures the similarity between a sketch and photo by their distance from the training set of sketch/photo pairs, and (3) by fusing the previous two methods. For both matching methods, the first step is to sample SIFT feature descriptors uniformly across all the sketch and photo images. In direct matching, we simply measure the distance of the SIFT descriptors between sketches and photos. In common representation matching, the distance between the descriptor vectors of the probe sketches and gallery photos at each local sample point is measured. This results in a vector of distances across the sketch or photo image to each member of the training basis. Further recognition improvements are shown by score level fusion of the two sketch matchers. Compared with published sketch to photo matching algorithms, experimental results demonstrate improved matching performances using the presented feature-based methods.

A reliable thermal face recognition system can enhance the national security applications such as prevention against terrorism, surveillance, monitoring and tracking, especially at nighttime. The system can be applied at airports, customs or high-alert facilities (e.g., nuclear power plant) for 24 hours a day. In this paper, we propose a novel face recognition approach utilizing thermal (long wave infrared) face images that can automatically identify a subject at both daytime and nighttime. With a properly acquired thermal image (as a query image) in monitoring zone, the following processes will be employed: normalization and denoising, face detection, face alignment, face masking, Gabor wavelet transform, face pattern words (FPWs) creation, face identification by similarity measure (Hamming distance). If eyeglasses are present on a subject's face, an eyeglasses mask will be automatically extracted from the querying face image, and then masked with all comparing FPWs (no more transforms). A high identification rate (97.44% with Top-1 match) has been achieved upon our preliminary face dataset (of 39 subjects) from the proposed approach regardless operating time and glasses-wearing condition.e

Human faces are smooth and symmetrical, making superquadrics a good choice for representation and normalization. We present a novel approach to parameterize 3D faces using the powerful superquadric model in combination with an Eigen decomposition which represents the finer features of faces. The superquadric fit also provides axes of symmetry that yield a normalized face coordinate space necessary for applying PCA. Results of fitting on our data set, of 2 scans each from 107 people, show reliable representation for yaw, pitch, and roll with average rotations of the order 10^-3 radians, about each axis. Parameterization can be used to partition the search space into smaller bins, thus effectively reducing the search space complexity for matching and recognition. We show that it is possible to create about 20-40 clusters with as few as 30 parameters. The accuracy of the clustering algorithm, in some cases, is as high as 90%. We believe this approach to indexing 3D faces is an interesting extension to existing literature.

A new face recognition algorithm has been proposed which is robust to variations in pose, expression, illumination and occlusions such as sunglasses. The algorithm is motivated by the Edit Distance used to determine the similarity between strings of one dimensional data such as DNA and text. The key to this approach is how to extend the concept of an Edit Distance on one-dimensional data to two-dimensional image data. The algorithm is based on mapping one image into another and using the characteristics of the mapping to determine a two-dimensional Pictorial-Edit Distance or P-Edit Distance. We show how the properties of the mapping are similar to insertion, deletion and substitution errors defined in an Edit Distance. This algorithm is particularly well suited for face recognition in uncontrolled environments such as stand-off and other surveillance applications. We will describe an entire system designed for face recognition at a distance including face detection, pose estimation, multi-sample fusion of video frames and identification. Here we describe how the algorithm is used for face recognition at a distance, present some initial results and describe future research directions.(

We report the development of a face recognition system which operates in the same way as humans in that it is capable of recognizing a number of people, while rejecting everybody else as strangers. While humans do it routinely, a particularly challenging aspect of the problem of open-world face recognition has been the question of rejecting previously unseen faces as unfamiliar. Our approach can handle previously unseen faces; it is based on identifying and enclosing the region(s) in the human face space which belong to the target person(s).

Automatic feature extraction in latent fingerprints is a challenging problem due to poor quality of most latents, such as unclear ridge structures, overlapped lines and letters, and overlapped fingerprints. We proposed a latent fingerprint enhancement algorithm which requires manually marked region of interest (ROI) and singular points. The core of the proposed enhancement algorithm is a novel orientation field estimation algorithm, which fits orientation field model to coarse orientation field estimated from skeleton outputted by a commercial fingerprint SDK. Experimental results on NIST SD27 latent fingerprint database indicate that by incorporating the proposed enhancement algorithm, the matching accuracy of the commercial matcher was significantly improved.

In some applications such as field stations, disaster situations or similar conditions, it is desirable to have a contactless, rugged means to collect fingerprint information. The approach described in this paper enables acceleration of the capture process by eliminating an otherwise system and finger cleanup procedure, minimizes the chance of the spread of disease or contaminations, and uses an innovative optical system able to provide rolled equivalent fingerprint information desirable for reliable 2D matching against existing databases. The approach described captures highresolution fingerprints and 3D information simultaneously using a single camera. Liquid crystal polarization rotators combined with birefringent elements provides the focus shift and a depth from focus algorithm extracts the 3D data. This imaging technique does not involve any moving parts, thus reducing cost and complexity of the system as well as increasing its robustness. Data collection is expected to take less than 100 milliseconds, capturing all four-finger images simultaneously to avoid sequencing errors. This paper describes the various options considered for contactless fingerprint capture, and why the particular approach was ultimately chosen.

Integrity of fingerprint data is essential to biometric and forensic applications. Accordingly, the FBI's Criminal Justice Information Services (CJIS) Division has sponsored development of software tools to facilitate quality control functions relative to maintaining its fingerprint data assets inherent to the Integrated Automated Fingerprint Identification System (IAFIS) and Next Generation Identification (NGI). This paper provides an introduction of two such tools. The first FBI-sponsored tool was developed by the National Institute of Standards and Technology (NIST) and examines and detects the spectral signature of the ridge-flow structure characteristic of friction ridge skin. The Spectral Image Validation/Verification (SIVV) utility differentiates fingerprints from non-fingerprints, including blank frames or segmentation failures erroneously included in data; provides a "first look" at image quality; and can identify anomalies in sample rates of scanned images. The SIVV utility might detect errors in individual 10-print fingerprints inaccurately segmented from the flat, multi-finger image acquired by one of the automated collection systems increasing in availability and usage. In such cases, the lost fingerprint can be recovered by re-segmentation from the now compressed multi-finger image record. The second FBI-sponsored tool, CropCoeff was developed by MITRE and thoroughly tested via NIST. CropCoeff enables cropping of the replacement single print directly from the compressed data file, thus avoiding decompression and recompression of images that might degrade fingerprint features necessary for matching.

We discuss the problem of preserving the privacy of a digital face image stored in a central database. In the proposed scheme, a private face image is dithered into two host face images such that it can be revealed only when both host images are simultaneously available; at the same time, the individual host images do not reveal the identity of the original image. In order to accomplish this, we appeal to the field of Visual Cryptography. Experimental results confirm the following: (a) the possibility of hiding a private face image in two unrelated host face images; (b) the successful matching of face images that are reconstructed by superimposing the host images; and (c) the inability of the host images, known as sheets, to reveal the identity of the secret face image.

Biometric systems usually do not possess a cryptographic level of security: it has been deemed impossible to perform a biometric authentication in the encrypted domain because of the natural variability of biometric samples and of the cryptographic intolerance even to a single bite error. Encrypted biometric data need to be decrypted on authentication, which creates privacy and security risks. On the other hand, the known solutions called "Biometric Encryption (BE)" or "Fuzzy Extractors" can be cracked by various attacks, for example, by running offline a database of images against the stored helper data in order to obtain a false match. In this paper, we present a novel approach which combines Biometric Encryption with classical Blum-Goldwasser cryptosystem. In the "Client - Service Provider (SP)" or in the "Client - Database - SP" architecture it is possible to keep the biometric data encrypted on all the stages of the storage and authentication, so that SP never has an access to unencrypted biometric data. It is shown that this approach is suitable for two of the most popular BE schemes, Fuzzy Commitment and Quantized Index Modulation (QIM). The approach has clear practical advantages over biometric systems using "homomorphic encryption". Future work will deal with the application of the proposed solution to one-to-many biometric systems.

Template protection techniques are used within biometric systems in order to protect the stored biometric template against privacy and security threats. A great portion of template protection techniques are based on extracting a key from or binding a key to a biometric sample. The achieved protection depends on the size of the key and its closeness to being random. In the literature it can be observed that there is a large variation on the reported key lengths at similar classification performance of the same template protection system, even when based on the same biometric modality and database. In this work we determine the analytical relationship between the system performance and the theoretical maximum key size given a biometric source modeled by parallel Gaussian channels. We consider the case where the source capacity is evenly distributed across all channels and the channels are independent. We also determine the effect of the parameters such as the source capacity, the number of enrolment and verification samples, and the operating point selection on the maximum key size. We show that a trade-off exists between the privacy protection of the biometric system and its convenience for its users.

Fuzzy vault is a practical and promising scheme, which can protect biometric templates and perform secure key management simultaneously. Aligning the query sample and the template sample in the encrypted domain remains a challenging task in the fingerprint-based fuzzy vault scheme. To some extent, all the existing fingerprint aligning methods in the encrypted domain have their own drawbacks, e.g., not enough alignment accuracy or information leakage because of publishing helper data. In this paper, a novel fingerprint aligning method is proposed, which integrates the fingerprint reference points and its neighboring region of interest(ROI) in a hierarchical manner. The concept of mutual information(MI) in the information theory is used to assess the coincidence extent of two fingerprints after being aligned. The novel alignment method is applied to fingerprint-based fuzzy vault implementation. Out of information leakage consideration, the orientation features of fingerprint minutiae are discarded and another distinguishing local feature, inter-minutiae ridge count, is used to replace the minutiae orientation in the implementation of fingerprint-based fuzzy vault. Experiment on FVC2002 DB2a is conducted to show the virtue of proposed alignment method and the promising performance of proposed fingerprint-based fuzzy vault implementation.

In this paper, we present a central limit theorem (CLT) for the estimation of a false match rate for a single matching system. The false match rate is often a significant factor in an evaluation of such a matching system. To achieve the main result here we utilize the covariance/correlation structure for matching proposed by Schuckers. Along with the main result we present an illustration of the methodology here on biometric authentication data from Ross and Jain. This illustration is from resampling match decisions on three different biometric modalities: hand geometry, fingerprint and facial recognition and shows that as the number of matching pairs grows the sampling distribution for an FMR approaches a Gaussian distribution. These results suggest that statistical inference for a FMR based upon a Gaussian distribution is appropriate.

It is not uncommon for contemporary biometric systems to have more than one match below the matching threshold, or to have two or more matches having close matching scores. This is especially true for those that store large quantities of identities and/or are applied to measure loosely constrained biometric traits, such as in identification from video or at a distance. Current biometric performance evaluation standards however are still largely based on measuring single-score statistics such as False Match, False Non-Match rates and the trade-off curves based thereon. Such methodology and reporting makes it impossible to investigate the risks and risk mitigation strategies associated with not having a unique identifying score. To address the issue, Canada Border Services Agency has developed a novel modality-agnostic multi-order performance analysis framework. The framework allows one to analyze the system performance at several levels of detail, by defining the traditional single-score-based metrics as Order-1 analysis, and introducing Order- 2 and Order-3 analysis to permit the investigation of the system reliability and the confidence of its recognition decisions. Implemented in a toolkit called C-BET (Comprehensive Biometrics Evaluation Toolkit), the framework has been applied in a recent examination of the state-of-the art iris recognition systems, the results of which are presented, and is now recommended to other agencies interested in testing and tuning the biometric systems.

Existing definitions for biometric testing and evaluation do not fully explain errors in a biometric system. This paper provides a definitional framework for the Human Biometric-Sensor Interaction (HBSI) model. This paper proposes six new definitions based around two classifications of presentations, erroneous and correct. The new terms are: defective interaction (DI), concealed interaction (CI), false interaction (FI), failure to detect (FTD), failure to extract (FTX), and successfully acquired samples (SAS). As with all definitions, the new terms require a modification to the general biometric model developed by Mansfield and Wayman [1].

To evaluate the performance of fingerprint-image matching algorithms on large datasets, a receiver operating characteristic (ROC) curve is applied. From the operational perspective, the true accept rate (TAR) of the genuine scores at a specified false accept rate (FAR) of the impostor scores and/or the equal error rate (EER) are often employed. Using the standard errors of these metrics computed using the nonparametric two-sample bootstrap based on our studies of bootstrap variability on large fingerprint datasets, the significance test is performed to determine whether the difference between the performance of one algorithm and a hypothesized value, or the difference between the performances of two algorithms where the correlation is taken into account is statistically significant. In the case that the alternative hypothesis is accepted, the sign of the difference is employed to determine which is better than the other. Examples are provided.

Iris recognition technology has the potential to broaden to include non-ideal imaging situations, as well as scale to national-level deployments. Hence, the study of population factors, subject intrinsics, and sensing contexts that can individually or collectively impact iris recognition performance assumes increased importance. This presentation summarizes recent research on a number of such "quality variables" and motivates the need for large data sets exhibiting a variety of such non-idealities to adequately characterize performance.

Low quality iris images such as blurry, low resolution images with poor illumination create a challenge for iris recognition systems. Therefore, an efficient enhancement of iris images are needed in challenging environments. We propose a new iris recognition algorithm for enhancement of normalized iris images. Our algorithm is based on the logarithmic image processing (LIP) image enhancement which is used as one of the 3 stages in the enhancement process. Methods are tested on the MBGC database to compare enrolled video iris images from 124 subjects with 220 pixels resolutions to query video portal images from 110 subjects with 120 pixels resolution. Results from processing challenging MBGC iris data show significant improvement in the performance of iris recognition algorithms in terms of equal error rates compared to the original (unenhanced images) and the other fast image enhancement methods.

Conventional computer systems authenticate users only at the initial log-in session, which can be the cause of a critical security flaw. To resolve this problem, systems need continuous user authentication methods that continuously monitor and authenticate users based on some biometric trait(s). We propose a new method for continuous user authentication based on a Webcam that monitors a logged in user's face and color of clothing. Our method can authenticate users regardless of their posture in front of the workstation (laptop or PC). Previous methods for continuous user authentication cannot authenticate users without biometric observation. To alleviate this requirement, our method uses color information of users' clothing as an enrollment template in addition to their face information. The system cannot pre-register the clothing color information because this information is not permanent. To deal with the problem, our system automatically registers this information every time the user logs in and then fuses it with the conventional (password) identification system. We report preliminary authentication results and future enhancements to the proposed system.

Biometrics, such as fingerprint, iris scan, and face recognition, offer methods for identifying individuals based on a unique physiological measurement. Recent studies indicate that a person's electrocardiogram (ECG) may also provide a unique biometric signature. Several methods for processing ECG data have appeared in the literature and most approaches rest on an initial detection and segmentation of the heartbeats. Various sources of noise, such as sensor noise, poor sensor placement, or muscle movements, can degrade the ECG signal and introduce errors into the heartbeat segmentation. This paper presents a screening technique for assessing the quality of each segmented heartbeat. Using this technique, a higher quality signal can be extracted to support the identification task. We demonstrate the benefits of this quality screening using a principal component technique known as eigenpulse. The analysis demonstrated the improvement in performance attributable to the quality screening.

Biometrics is described as the science of identifying people based on physical characteristics such as their fingerprints, facial features, hand geometry, iris patterns, palm prints, or speech recognition. Notably, all of these physical characteristics are visible or detectable from the exterior of the body. These external characteristics can be lifted, photographed, copied or recorded for unauthorized access to a biometric system. Individual humans are unique internally, however, just as they are unique externally. New biometric modalities have been developed which identify people based on their unique internal characteristics. For example, "Boneprints^TM" use acoustic fields to scan the unique bone density pattern of a thumb pressed on a small acoustic sensor. Thanks to advances in piezoelectric materials the acoustic sensor can be placed in virtually any device such as a steering wheel, door handle, or keyboard. Similarly, "Imp-PrintsTM" measure the electrical impedance patterns of a hand to identify or verify a person's identity. Small impedance sensors can be easily embedded in devices such as smart cards, handles, or wall mounts. These internal biometric modalities rely on physical characteristics which are not visible or photographable, providing an added level of security. In addition, both the acoustic and impedance methods can be combined with physiologic measurements such as acoustic Doppler or impedance plethysmography, respectively. Added verification that the biometric pattern came from a living person can be obtained. These new biometric modalities have the potential to allay user concerns over protection of privacy, while providing a higher level of security.*

We present a prototype video tracking and person categorization system that uses face and person soft biometric features to tag people while tracking them in multiple camera views. Our approach takes advantage of temporal aspect of video by extracting and accumulating feasible soft biometric features for each person in every frame to build a dynamic soft biometric feature list for each tracked person in surveillance videos. We developed algorithms for extracting face soft biometric features to achieve gender and ethnicity classification and session soft biometric features to aid in camera hand-off in surveillance videos with low resolution and uncontrolled illumination. To train and test our face soft biometry algorithms, we collected over 1500 face images from both genders and three ethnicity groups with various sizes, poses and illumination. These soft biometric feature extractors and classifiers are implemented on our existing video content extraction platform to enhance video surveillance tasks. Our algorithms achieved promising results for gender and ethnicity classification, and tracked person re-identification for camera hand-off on low to good quality surveillance and broadcast videos. By utilizing the proposed system, a high level description of extracted person's soft biometric data can be stored to use later for different purposes, such as to provide categorical information of people, to create database partitions to accelerate searches in responding to user queries, and to track people between cameras.

The need for an automated surveillance system is pronounced at night when the capability of the human eye to detect anomalies is reduced. While there have been significant efforts in the classification of individuals using human metrology and gait, the majority of research assumes a day-time environment. The aim of this study is to move beyond traditional image acquisition modalities and explore the issues of object detection and human identification at night. To address these issues, a spatiotemporal gait curve that captures the shape dynamics of a moving human silhouette is employed. Initially proposed by Wang et al., this representation of the gait is expanded to incorporate modules for individual classification, backpack detection, and silhouette restoration. Evaluation of these algorithms is conducted on the CASIA Night Gait Database, which includes 10 video sequences for each of 153 unique subjects. The video sequences were captured using a low resolution thermal camera. Matching performance of the proposed algorithms is evaluated using a nearest neighbor classifier. The outcome of this work is an efficient algorithm for backpack detection and human identification, and a basis for further study in silhouette enhancement.

In this paper we present a prototype for an automated deception detection system. Similar to polygraph examinations, we attempt to take advantage of the theory that false answers will produce distinctive measurements in certain physiological manifestations. We investigate the role of dynamic eye-based features such as eye closure/blinking and lateral movements of the iris in detecting deceit. The features are recorded both when the test subjects are having non-threatening conversations as well as when they are being interrogated about a crime they might have committed. The rates of the behavioral changes are blindly clustered into two groups. Examining the clusters and their characteristics, we observe that the dynamic features selected for deception detection show promising results with an overall deceptive/non-deceptive prediction rate of 71.43% from a study consisting of 28 subjects.

In forensic authentication, one aims to identify the perpetrator among a series of suspects or distractors. A fundamental problem in any recognition system that aims for identification of subjects in a natural scene is the lack of constrains on viewing and imaging conditions. In forensic applications, identification proves even more challenging, since most surveillance footage is of abysmal quality. In this context, robust methods for pose estimation are paramount. In this paper we will therefore present a new pose estimation strategy for very low quality footage. Our approach uses 3D-2D registration of a textured 3D face model with the surveillance image to obtain accurate far field pose alignment. Starting from an inaccurate initial estimate, the technique uses novel similarity measures based on the monogenic signal to guide a pose optimization process. We will illustrate the descriptive strength of the introduced similarity measures by using them directly as a recognition metric. Through validation, using both real and synthetic surveillance footage, our pose estimation method is shown to be accurate, and robust to lighting changes and image degradation.

The increase in twin births has created a requirement for biometric systems to accurately determine the identity of a person who has an identical twin. The discriminability of some of the identical twin biometric traits, such as fingerprints, iris, and palmprints, is supported by anatomy and the formation process of the biometric characteristic, which state they are different even in identical twins due to a number of random factors during the gestation period. For the first time, we collected multiple biometric traits (fingerprint, face, and iris) of 66 families of twins, and we performed unimodal and multimodal matching experiments to assess the ability of biometric systems in distinguishing identical twins. Our experiments show that unimodal finger biometric systems can distinguish two different persons who are not identical twins better than they can distinguish identical twins; this difference is much larger in the face biometric system and it is not significant in the iris biometric system. Multimodal biometric systems that combine different units of the same biometric modality (e.g. multiple fingerprints or left and right irises.) show the best performance among all the unimodal and multimodal biometric systems, achieving an almost perfect separation between genuine and impostor distributions.

In this work we present a multibiometric face recognition framework based on combining information from 2D with 3D facial features. The 3D biometrics channel is protected by a privacy enhancing technology, which uses error correcting codes and cryptographic primitives to safeguard the privacy of the users of the biometric system at the same time enabling accurate matching through fusion with 2D. Experiments are conducted to compare the matching performance of such multibiometric systems with the individual biometric channels working alone and with unprotected multibiometric systems. The results show that the proposed hybrid system incorporating template protection, match and in some cases exceed the performance of corresponding unprotected equivalents, in addition to offering the additional privacy protection.

While fusion can be accomplished at multiple levels in a multibiometric system, score level fusion is commonly used as it offers a good trade-off between fusion complexity and data availability. However, missing scores affect the implementation of several biometric fusion rules. While there are several techniques for handling missing data, the imputation scheme - which replaces missing values with predicted values - is preferred since this scheme can be followed by a standard fusion scheme designed for complete data. This paper compares the performance of three imputation methods: Imputation via Maximum Likelihood Estimation (MLE), Multiple Imputation (MI) and Random Draw Imputation through Gaussian Mixture Model estimation (RD GMM). A novel method called Hot-deck GMM is also introduced and exhibits markedly better performance than the other methods because of its ability to preserve the local structure of the score distribution. Experiments on the MSU dataset indicate the robustness of the schemes in handling missing scores at various missing data rates.

Active learning methods have gained popularity to reduce human effort in annotating examples in order to train a classifier. When faced with large amounts of data, the active learning algorithm automatically selects appropriate data samples that are most relevant to train the classifier. Typical active learning approaches select one data instance (one face image, for example) in one iteration of the algorithm, and the classifier is trained with the selected data instances, one-by-one. Instead, there have been very recent efforts in active learning to select a batch of examples for labeling at each instant rather than selecting a single example and updating the hypothesis. In this work, a novel batch mode active learning scheme based on numerical optimization of an appropriate function has been applied to the biometric recognition problem. In problems such as face recognition, real-world data is often generated in batches, such as frames of video in a capture session. In such scenarios, selecting the most appropriate data instances from these batches (which usually have a high redundancy) to train a classifier is a significant challenge. In this work, the instance selection is formulated as a mathematical optimization problem and the framework is extended to handle learning from multiple sources of information. The results obtained on the widely used NIST Multiple Biometric Grand Challenge (MBGC) and VidTIMIT biometric datasets corroborate the potential of this method in being used for real-world biometric recognition problems, when there are large amounts of data.

We assess the impact of the H.264 video codec on the match performance of automated face recognition in surveillance and mobile video applications. A set of two hundred access control (90 pixel inter-pupilary distance) and distance surveillance (45 pixel inter-pupilary distance) videos taken under non-ideal imaging and facial recognition (e.g., pose, illumination, and expression) conditions were matched using two commercial face recognition engines in the studies. The first study evaluated automated face recognition performance on access control and distance surveillance videos at CIF and VGA resolutions using the H.264 baseline profile at nine bitrates rates ranging from 8kbs to 2048kbs. In our experiments, video signals were able to be compressed up to 128kbs before a significant drop face recognition performance occurred. The second study evaluated automated face recognition on mobile devices at QCIF, iPhone, and Android resolutions for each of the H.264 PDA profiles. Rank one match performance, cumulative match scores, and failure to enroll rates are reported.

As a new biometrics authentication technology, ear recognition remains many unresolved problems, one of them is the occlusion problem. This paper deals with ear recognition with partially occluded ear images. Firstly, the whole 2D image is separated to sub-windows. Then, Neighborhood Preserving Embedding is used for feature extraction on each subwindow, and we select the most discriminative sub-windows according to the recognition rate. Thirdly, a multi-matcher fusion approach is used for recognition with partially occluded images. Experiments on the USTB ear image database have illustrated that using only few sub-window can represent the most meaningful region of the ear, and the multimatcher model gets higher recognition rate than using the whole image for recognition.