The growing deployment of biometrics as a proof of identity has generated a great deal of research into biometrics in recent years, and widened the scope of investigations beyond improving accuracy into mechanisms to deal with serious concerns raised about security and privacy due to the potential misuse. of the collected biometric data along with possible attacks on biometric systems. The focus on improving performance of biometric authentication has been more on multi-modal and multi-factor biometric authentication in conjunction with designing recognition techniques to mitigate the adverse effect of variations in recording conditions. Some of these approaches together with the emerging developments of cancellable biometrics and biometric cryptosystems have been used as mechanisms to enhance security and privacy of biometric systems. This thesis is designed to deal with these complimentary and closely related issues through investigations that aim at understanding the impact of varying biometric sample recording conditions on the discriminating information content (entropy) of these samples, and to use the gained knowledge to (1) design adaptive techniques for improved performance of biometric authentication, and (2) propose and test a framework for a proper evaluation of security of all factors/components involved in biometric keys and multi-factor biometric authentication. The first part of this thesis consists of a set of theoretical and empirical investigations designed to evaluate and analyse the effect of emerging developments in biometrics systems, with a focus on those related to biometric entropy and multi-factor authentication. The analysis of different biometric entropy measures, proposed in the literature, reveals that variations in biometric sample quality lead to variations in the correlation between biometric entropy values calculated using any of the known measures and the accuracy of the biometric recognition. Furthermore, analysis of the spatial distribution of entropy values in face images reveals a non-uniform distribution. The widely expected inherent individual differences in biometric features entropy will also be confirmed. Moreover, we uncover a myth reported in the literature about near perfect accuracy of certain quality-based adaptive recognition schemes.

This thesis provides details relating to the developments of work by the author in partial fulfilment of degree of Doctor of Philosophy within the University of Sunderland's laboratories. The author discusses efficient alternative methods that are employed to detect two dimensional symbols embodied within an image. The symbols of particular interest represent telecom network components from a paper map. British Telecom store network component maps in scanned digital format taken from paper maps which have been subjected to updates over many years with respect to new and modified equipment. The aim of this project is to create a 'netlist' of components and their circuit connection. This 'netlist' offers a descriptive circuit topology that can be interfaced and employed in the creation of network schematic for inclusion into other applications such as planning software. The author has adopted an approach to follow fundamental human qualitative methods of image recognition within the novel design of quantitative machine recognition methods. Principles of evolution are employed within the development of an application which include a number of novel algorithms based on soft computing. Colour frequency analysis and morphologic processing are also employed as methods to implement recognition evaluation. These algorithms are encompassed within an Evolutionary Image Recognition Algorithm (EIRA). A series of experiments have been carried out to determine the efficiency of the employed methods for the symbol extraction and results have been obtained that show the approach adopted by the author provides a rapid symbol recognition system.

Laughter is clearly an audiovisual event, consisting of the laughter vocalisation and involving facial activity around the mouth. Past research on automatic laughter classification has focused mainly on audio-based approaches. In this thesis we integrate the information from audio and video channels and show that this fusion may lead to improved performance over unimodal approaches. We investigated different types of audiovisual fusion, temporal modelling and feature sets in order to find the best combination. A novel approach to combine audio and visual information based on prediction is also proposed, which explicitly models spatial and temporal relationship between audio and visual features. Experiments are presented both on matched training and test conditions, using subject-independent cross validation in one database, and unmatched conditions using 6 databases. This presents a challenging situation which is rarely addressed in the literature. Comparison of the different fusion approaches is performed on these databases, confirming that the prediction-based method proposed usually performs better than standard fusion methods. The lack of suitable data is a major obstacle in studying laughter so we introduce a new publicly available audiovisual database suitable for studying laughter. It contains 22 subjects which were recorded while watching stimulus material, by two microphones, a video camera and a thermal camera. An analysis of the errors of the audio, video and audiovisual classifiers is also performed in terms of gender, language, laughter types and noise levels in order to get an insight of when visual information helps. Finally, results on the first attempt to discriminate two types of laughter, voiced and unvoiced, in an audiovisual way are presented. Overall, it is demonstrated that in most cases the addition of visual information to audio leads to improved performance in laughter-vs-speech discrimination and audiovisual fusion is really beneficial as the audio noise levels increase.

Visual search is· a ubiquitous task in everyday life. Simple tasks, from finding cars keys to reaching for a door handle, all require an element of visual search. In some critical situations, such as in security surveillance or for Search and Rescue situations, there is an advantage in augmenting the human search capability with computer assistance. The aim of this thesis is to investigate different computational models of human visual perception in the context of natural image search .. Using two psychophysical experiments to investigate human visual search behaviour, this thesis demonstrates similarities and differences between search tasks using completely natural scenes and the previous literature of artificially constructed search scenes. The human experimental results are compared to the results from several computational models of scene content that use the analysis of either the clutter content of a scene, salience of objects in the scene, or the relevance to the target of locations in a scene. Additionally a new model of visual attention is developed, using the principle of a relevancy heat map to identify regions of a scene that match a specific target. The results suggest that the type of model that best predicts visual search behaviour depends upon the nature of the search task. Models that account for the details of the target through a principle of relevance are better for matching target-present behaviour and that models that use simple saliency of the, features better predict fixations in target-absent searches.

Recent interest in the preservation of our heritage has brought about increased archival research, which has made a considerable number of historical documents available in digital format. However, the analysis of these documents still greatly depends on manually intensive work by domain experts. Early music manuscripts are one of the most complicated sources as they require extensive knowledge of domain experts. Although optical music recognition has been actively investigated, it has; not been applied to music manuscripts from a historical musicological perspective. The principal aim of this work is to reveal the potential of music manuscripts as a source of data mining by integrating historical musicologists' knowledge. Particular attention is paid to the paleographical aspects of music manuscripts. Image processing is used to extract geometric features from music manuscripts and statistical analysis is conducted. The proposed methods are validated by exploring case studies in Bach source studies, the results of which suggest that there is a strong potential for them to become a road map not only for musicological research but also other empirical research. Moreover, the results of the proposed research may, also serve as a prototype for next-generation data mining, which can not only use the data-driven information but also the experts' background knowledge in highly professional subjects

No description available.

Among the biometric systems that have been developed so far, iris recognition systems have emerged as being one of the most reliable. In iris recognition, most of the ,: research was conducted on operation under near infrared illumination. For unconstrained scenarios of iris recognition systems. the iris images are captured under visible light spectrum and therefore incorporate various types of imperfections. In this thesis the merits of fusing information from various sources for improving the state of the art accuracies of colour iris recognition systems is evaluated. An investigation of how fundamentally different fusion strategies can increase the degree of choice available in achieving certain performance criteria is conducted. Initially, simple fusion mechanisms are employed to increase the accuracy of an iris recognition system and then more complex fusion architectures are elaborated to further enhance the biometric system's accuracy. In particular, the design process of the iris recognition system with reduced constraints is carried out using three different fusion approaches: multi-algorithmic, texture and colour fusion and multiple classifier systems. In the first approach, one novel iris feature extraction methodology is proposed and a multi-algorithmic iris recognition system using score fusion, composed. of 3 individual systems, is benchmarked. In the texture and colour fusion approach, the advantages of fusing information from the iris texture with data extracted from the eye colour are illustrated. Finally, the multiple classifier systems approach investigates how the robustness and practicability of an iris recognition system operating on visible spectrum images can be enhanced by training individual classifiers on different iris features. Besides the various fusion techniques explored, an iris segmentation' algorithm is proposed and a methodology for finding which colour channels from a colour space reveal the most discriminant information from the iris texture is introduced. The contributions presented in this thesis indicate that iris recognition systems that operate on visible spectrum images can be designed to operate with an accuracy required by a particular application scenario. Also, the iris recognition systems developed in the present study are suitable for mobile and embedded implementations.

In this thesis the problem of estimating the 2-D articulated pose, or configuration of a person in unconstrained images such as consumer photographs is addressed. Contributions are split among three major chapters. In previous work the Pictorial Structure Model approach has proven particularly successful. and is appealing because of its moderate computational cost. However, the accuracy of resulting pose estimates has been limited by the use of simple representations of limb appearance. In this thesis strong discriminatively trained limb detectors combining gradient and colour segmentation cues are proposed. The approach improves significantly on the "iterative image parsing" method which was the state-of-the-art at the time, and shows significant promise for combination with other models of pose and appearance. In the second pan of this thesis higher fidelity models of pose and appearance are proposed. The aim is to tackle extremely challenging properties of the human pose estimation task arising from variation in pose, anatomy, clothing. and imaging conditions. Current methods use simple models of body part appearance and plausible configurations due to limitations of available training data and constraints on computational expense. It is shown that such models severely limit accuracy. A new annotated database of challenging consumer images is introduced, an order of magnitude larger than currently available datasets. This larger amount of data allows partitioning of the pose space and the learning of multiple, clustered Pictorial Structure Models. A relative improvement in accuracy of over 50% is achieved compared to the standard, single model approach. In the final part of this thesis the clustered Pictorial Structure Model framework is extended to handle much larger quantities of training data. Furthermore it is shown how to utilise Amazon Mechanical Turk and a latent annotation update scheme to achieve high quality annotations at low cost. A significant increase in pose estimation accuracy is presented, while the computational expense of the framework is improved by a factor of

No description available.

No description available.