The research aimed to explore the potential occurrence of the serial position effect and the own-race bias in facial recognition, and to explore whether these two socio-cognitive psychological phenomena had any influence on each other. Specifically, the researcher suggested that other-race facial recognition will show diminished U-type serial position curves in comparison to own-race facial recognition U-type serial position curves. This was done through a quasi-experimental design, testing 48 participants from an environmental consulting and ground engineering firm in Midrand, Johannesburg. Twelve (12) sets of slides showing either 5 black or 5 white faces were presented to participants, the sequence of faces was randomised and then displayed again to participants. Participants had to identify the original position in which the face was displayed (forcing a serial reconstruction task). Results yielded a U-type serial position curves for overall recognition, with a statistically significant effect for own-race bias effect. Specific interactions indicated that recognition for own-race facial stimuli exhibit clear serial position effect trends, whilst recognition of other-race facial stimuli recognition show increased recognition for the first, middle and last faces in a set. The researcher suggests that the results within this study could be attributed to the effect of an attentional primacy gradient within the Serial Information Processing model. However, further studies are required to eliminate numerous other confounding factors which may have played a role in the study. The results of this research have implications for the judicial system, in which false eyewitness identifications have profound consequences. / Dissertation (MA)--University of Pretoria, 2010. / Psychology / unrestricted

This dissertation develops an image processing framework with unique feature extraction and similarity measurements for human face recognition in the mid-wave infrared portion of the electromagnetic spectrum. The goal is to design specialized algorithms that would extract vasculature information, create a thermal facial signature and identify the individual. The objective is to use such findings in support of a biometrics system for human identification with a high degree of accuracy and a high degree of reliability. This last assertion is due to the minimal to no risk for potential alteration of the intrinsic physiological characteristics seen through thermal imaging. Thermal facial signature authentication is fully integrated and consolidates the main and critical steps of feature extraction, registration, matching through similarity measures, and validation through the principal component analysis. Feature extraction was accomplished by first registering the images to a reference image using the functional MRI of the Brain’s (FMRIB’s) Linear Image Registration Tool (FLIRT) modified to suit thermal images. This was followed by segmentation of the facial region using an advanced localized contouring algorithm applied on anisotropically diffused thermal images. Thermal feature extraction from facial images was attained by performing morphological operations such as opening and top-hat segmentation to yield thermal signatures for each subject. Four thermal images taken over a period of six months were used to generate a thermal signature template for each subject to contain only the most prevalent and consistent features. Finally a similarity measure technique was used to match images to the signature templates and the Principal Component Analysis (PCA) was used to validating the results of the matching process. Thirteen subjects were used for testing the developed technique on an in-house thermal imaging system. The matching using the similarity measures showed 88% accuracy in case of skeletonized feature signatures and 90% accuracy for anisotropically diffused feature signatures. The highly accurate results obtained in the matching process along with the generalized design process clearly demonstrate the ability of the developed thermal infrared system to be used on other thermal imaging based systems and related databases.

With the usage of biometrics becoming commonly used in a variety of applications, keeping those biometrics private and secure is an important issue. Indeed, the convenience of using biometrics for authentication is counteracted by the fact that they cannot easily be modified or changed. This can have dire consequences to a person if their biometrics are leaked. In the past decades, various techniques have been proposed to solve this problem. Such techniques range from using and storing randomized templates, using homomorphic encryption, or using biometric encryption techniques such as fuzzy extractors. Fuzzy extractors are a construction that allows the extraction of cryptographic keys from noisy data like biometrics. The key can then be rebuilt from some helper data and another biometric, provided that it is similar enough to the biometrics used to generate the key. This can be achieved through various approaches like the use of a quantizer or an error correcting code. In this thesis, we consider specifically fuzzy extractors for facial images. The first part of this thesis focuses on improving the security, privacy and performance of the extractor for faces first proposed by Sutcu et al. Our improvements make their construction more resistant to partial and total leaks of secure information, as well as improve the performance in a biometric authentication setting. The second part looks at using low density lattice codes (LDLC) as a quantizer in the fuzzy extractor, instead of using component based quantization. Although LDLC have been proposed as a quantizer for a general fuzzy extractor, they have yet to be used or tested for continuous biometrics like face images. We present a construction for a fuzzy extractor scheme using LDLC and we analyze its performance on a publicly available data set of images. Using an LDLC quantizer on this data set has lower accuracy than the improved scheme from the first part of this thesis. On the other hand, the LDLC scheme performs better when the inputs have additive white Gaussian noise (AWGN), as we show through simulated data. As such, we expect it to perform well in general on data and biometrics with variance akin to a AWGN channel.

La Reconnaissance automatique des personnes est devenue un sujet de plus en plus important avec l'augmentation constante des besoins en sécurité. De nombreux systèmes biométriques existent. Ils utilisent différentes caractéristiques humaines. Parmi tous les traits biométriques, la reconnaissance faciale inclut des aspects positifs en termes d'accessibilité et de fiabilité. Dans cette thèse, deux défis en reconnaissance faciales sont étudiés. Le premier est le leurrage. Le leurrage en reconnaissance faciale est présenté. Des contre-mesures permettant d'améliorer les systèmes actuels sont proposés. A cet effet, les attaques basées sur des photographies 2D ou des masques 3D sont analysées. Le second défi exploré dans cette thèse est lié aux variations dues à des altérations du visage (i.e. chirurgie plastique), maquillage et accessoires pour le visage (e.g. occultations par la présence de lunettes). L'impact de ces variations en reconnaissance de visage est étudiée séparément. Ensuite, des techniques robustes contre les variations de camouflage sont proposées. / Human recognition has become an important topic as the need and investments for security applications grow continuously. Numerous biometric systems exist which utilize various human characteristics. Among all biometrics traits, face recognition is advantageous in terms of accessibility and reliability. In the thesis, two challenges in face recognition are analyzed. The first one is face spoofing. Spoofing in face recognition is explained together with the countermeasure techniques that are proposed for the protection of face recognition systems against spoofing attacks. For this purpose, both 2D photograph and 3D mask attacks are analyzed. The second challenge explored in the thesis is disguise variations, which are due to facial alterations, facial makeup and facial accessories (occlusions). The impact of these disguise variations on face recognition is explored, separately. Then, techniques which are robust against disguise variations are proposed.

Facial recognition systems pose numerous ethical challenges around privacy, racial and gender bias, and accuracy, yet little guidance is available for designers and developers. We explore solutions to these challenges in a four-phase design process to create Civil War Twin (CWT), an educational web-based application where users can discover their lookalikes from the American Civil War era (1861-65) while learning more about facial recognition and history. Through this design process, we synthesize industry guidelines, consult with scholars of history, gender, and race, evaluate CWT in feedback sessions with diverse prospective users, and conduct a usability study with crowd workers. We iteratively formulate design goals to incorporate transparency, inclusivity, speculative design, and empathy into our application. We found that users' perceived learning about the strengths and limitations of facial recognition and Civil War history improved after using CWT, and that our design successfully met users' ethical standards. We also discuss how our ethical design process can be applied to future facial recognition applications. / Master of Science / Facial recognition systems, such as those used in cities, smartphone application and airports, pose numerous ethical challenges around privacy, racial and gender bias, and accuracy. Little guidance is available for designers and developers to create ethical facial recognition systems. We explore solutions to these ethical challenges of creating facial recognition systems in a four-phase design process to create Civil War Twin (CWT), an educational web-based application where users can discover their lookalikes from the American Civil War era (1861-65) while learning more about facial recognition and history. CWT allows users to upload a selfie, select search preferences (e.g., military service, gender, ethnicity), and use facial recognition to discover their "Civil War twins" (i.e., photographs of people from the American Civil War era who look like them). Through this design process, we synthesize industry guidelines, consult with scholars of history, gender, and race, evaluate CWT in feedback sessions with diverse prospective users, and conduct a usability study. We iteratively formulate design goals to incorporate transparency, inclusivity, critical thinking, and empathy into our application. We found that users' perceived learning about the strengths and limitations of facial recognition and Civil War history improved after using CWT, and that our design successfully met users' ethical standards. We also discuss how our ethical design process can be applied to future facial recognition applications.

Yes / This study aims to investigate the role of artificial intelligence (AI) driven facial recognition to enhance a value proposition by influencing different areas of services in the travel and tourism industry. We adopted semi-structured interviews to derive insights from 26 respondents. Thematic analysis reveals the development of four main themes (personalization, data-driven service offering, security and safety, and seamless payments). Further, we mapped the impact of AI- driven facial recognition to enhance value and experience for corporate guests. Findings indicate that AI-based facial recognition can facilitate the travel and tourism industry in understanding travelers’ needs, optimization of service offers, and value-based services, whereas data-driven services can be realized in the form of customized trip planning, email, and calendar integration, and quick bill summarization. This contributes to strengthening the tourism literature through the lens of organizational information processing theory. / Area of Excellence AI, Data Science & Business at NEOMA Business School, France under the fund number 416005

The present study investigated whether spatial ability is correlated with the ability to accurately recognize faces. A samp~e pf 36 undergraduates were recruited for this study. Participants completed two measures of spatial ability: The Purdue Visualization of Rotations Test as a measure of mental rotation and the Object Location Memory Test as a measure of memory for the location of objects in relation to one another. Facial recognition was measured usipg the Cambridge Face Metnory Test, which measures the recognition of faces in both upright and inverted positions. As predicted, a significant relationship was found between mental rotation and inverted facial recognition, r = .33,p < .05. There was not a significant relationship between object location memory and upright facial recognition, r=.07,p > .05. Interestingly, upright facial recognition was more closely associated with mental rotation, though the relationship did not reach statistical significance r = .24,p > .05. There was not a significant relationship between overall spatial ability and overall facial recognition, r = .17,p > .05. The traditional gender differences were found oh spatial ability such that males outperformed females on mental rotation (Males: M= 12.73, SD= 3.93; Females: M= 9.32, SD= 4.11) and females outperformed males on object location memory (Females: M= 18.80, SD= 5.53; Males: M= 14.09,_ SD= 8.19). A significant gender difference on facial recognition such that females outperform males was not found, which contradicts findings from past J research (McBain, Norton, & Chen, 2009). These results suggest that mental rotation is an important factor in human facial recognition.

The focus of this dissertation is a methodology that enables computer systems to classify different up-front images of human faces as belonging to one of the individuals to which the system has been exposed previously. The images can present variance in size, location of the face, orientation, facial expressions, and overall illumination. The approach to the problem taken in this dissertation can be classified as analytic as the shapes of individual features of human faces are examined separately, as opposed to holistic approaches to face recognition. The outline of the features is used to construct signature functions. These functions are then magnitude-, period-, and phase-normalized to form a translation-, size-, and rotation-invariant representation of the features. Vectors of a limited number of the Fourier decomposition coefficients of these functions are taken to form the feature vectors representing the features in the corresponding vector space. With this approach no computation is necessary to enforce the translational, size, and rotational invariance at the stage of recognition thus reducing the problem of recognition to the k-dimensional clustering problem. A recognizer is specified that can reliably classify the vectors of the feature space into object classes. The recognizer made use of the following principle: a trial vector is classified into a class with the greatest number of closest vectors (in the sense of the Euclidean distance) among all vectors representing the same feature in the database of known individuals. A system based on this methodology is implemented and tried on a set of 50 pictures of 10 individuals (5 pictures per individual). The recognition rate is comparable to that of most recent results in the area of face recognition. The methodology presented in this dissertation is also applicable to any problem of pattern recognition where patterns can be represented as a collection of black shapes on the white background.

Facial recognition is an increasingly popular security measure in scenarios with low computational power, such as phones and Raspberry Pi’s. There are many facial recognition techniques available. The aim is to compare three such techniques in both performance and time metrics. An experiment was conducted to compare the facial recognition techniques Convolutional Neural Network (CNN), Eigenface with the classifiers K-Nearest Neighbors (KNN) and support vector machines (SVM) and Fisherface with the classifiers KNN and SVM under the same conditions with a limited version of the LFW dataset. The Python libraries scikit-learn and OpenCV as well as the CNN implementation FaceNet were used. The results show that the CNN implementation of FaceNet is the best technique in all metrics except for prediction time. FaceNet achieved an F-score of 100% while the OpenCV implementation of Eigenface using SVM scored the worst at 15.5%. The technique with the lowest prediction time was the scikit-learn implementation of Fisherface with SVM.

Faces, along with the personal identities behind them, are effective elements in organizing a collection of consumer photos, as they represent who was involved. However, the accurate discrimination and subsequent recognition of face appearances is still very challenging. This can be attributed to the fact that faces are usually neither perfectly lit nor captured, particularly in the uncontrolled environments of consumer photos. Unlike, for instance, passport photos that only show faces stripped of their surroundings, Consumer Photo Collections contain a vast amount of meaningful context. For example, consecutively shot photos often correlate in time, location or scene. Further information can also be provided by the people appearing in photos, such as their demographics (ages and gender are often easier to surmise than identities), clothing, or the social relationships among co-occurring people. Motivated by this ubiquitous context, we propose and research people recognition approaches that consider contextual information within photos, as well as across entire photo collections. Our aim of leveraging additional contextual information (as opposed to only considering faces) is to improve recognition performance. However, instead of requiring users to explicitly label specific pieces of contextual information, we wish to implicitly learn and draw from the seemingly coherent content that exists inherently across an entire photo collection. Moreover, unlike conventional approaches that usually predict the identity of only one person’s appearance at a time, we lay out a semi-supervised approach to jointly recognize multiple peoples’ appearances across an entire photo collection simultaneously. As such, our aim is to find the overall best recognition solution. To make context-based joint recognition of people feasible, we research a sparse but efficient graph-based approach that builds on Markov Networks and utilizes distance-based face description methods. We show how to exploit the following specific contextual cues: time, social semantics, body appearances (clothing), gender, scene and ambiguous captions. We also show how to leverage crowd-sourced gamified feedback to iteratively improve recognition performance. Experiments on several datasets demonstrate and validate the effectiveness of our semisupervised graph-based recognition approach compared to conventional approaches.