Accent is cited as an issue for speech recognition systems. Our experiments showed that the ASR word error rate is up to seven times greater for accented speech compared with standard British English. The main objective of this research is to develop Automatic Speech Recognition (ASR) techniques that are robust to accent variation. We applied different acoustic modelling techniques to compensate for the effects of regional accents on the ASR performance. For conventional GMM-HMM based ASR systems, we showed that using a small amount of data from a test speaker to choose an accent dependent model using an accent identification system, or building a model using the data from N neighbouring speakers in AID space, will result in superior performance compared to that obtained with unsupervised or supervised speaker adaptation. In addition we showed that using a DNN-HMM rather than a GMM-HMM based acoustic model would improve the recognition accuracy considerably. Even if we apply two stages of accent followed by speaker adaptation to the GMM-HMM baseline system, the GMM-HMM based system will not outperform the baseline DNN-HMM based system. For more contemporary DNN-HMM based ASR systems we investigated how adding different types of accented data to the training set can provide better recognition accuracy on accented speech. Finally, we proposed a new approach for visualisation of the AID feature space. This is helpful in analysing the AID recognition accuracies and analysing AID confusion matrices.

The aim of semantic image understanding is to reveal the semantic meaning behind the image pixel. This thesis investigates problems related to semantic image understanding, and have made the following contributions. Our first contribution is to propose the usage of histogram matching in Multiple Kernel Learning. We treat the two-dimensional kernel matrix as an image and transfer the histogram matching algorithm in image processing to kernel matrix. Experiments on various computer vision and machine learning datasets have shown that our method can always boost the performance of state of the art MKL methods. Our second contribution is to advocate the segment-then-recognize strategy in pixel-level semantic image understanding. We have developed a new framework which tries to integrate semantic segmentation with low-level segmentation for proposing object consistent regions. We have also developed a novel method trying to integrate semantic segmentation with interactive segmentation. We found this segment-then-recognize strategy also works well on medical image data, where we designed a novel polar space random field model for proposing gland-like regions. In the realm of image-level semantic image understanding, our contribution is a novel way to utilize the random forest. Most of the previous works utilizing random forest store the posterior probabilities at each leaf node, and each random tree in the random forest is considered to be independent from each other. In contrast, we store the training samples instead of the posterior probabilities at each leaf node. We consider the random forest as a whole and propose the concept of semantic nearest neighbor and semantic similarity measure. Based on these two concepts, we devise novel methods for image annotation and image retrieval tasks.

This thesis is concerned with liveness detection for biometric systems and in particular for face recognition systems. Biometric systems are well studied and have the potential to provide satisfactory solutions for a variety of applications. However, presentation attacks (spoofng), where an attempt is made at subverting them system by making a deliberate presentation at the sensor is a serious challenge to their use in unattended applications. Liveness detection techniques can help with protecting biometric systems from attacks made through the presentation of artefacts and recordings at the sensor. In this work novel techniques for liveness detection are presented using gaze information. The notion of natural gaze stability is introduced and used to develop a number of novel features that rely on directing the gaze of the user and establishing its behaviour. These features are then used to develop systems for detecting spoofng attempts. The attack scenarios considered in this work include the use of hand held photos and photo masks as well as video reply to subvert the system. The proposed features and systems based on them were evaluated extensively using data captured from genuine and fake attempts. The results of the evaluations indicate that gaze-based features can be used to discriminate between genuine and imposter. Combining features through feature selection and score fusion substantially improved the performance of the proposed features.

The Biometric systems are commonly used as a fundamental tool by both government and private sector organizations to allow restricted access to sensitive areas, to identify the criminals by the police and to authenticate the identification of individuals requesting to access to certain personal and confidential services. The applications of these identification tools have created issues of security and privacy relating to personal, commercial and government identities. Over the last decade, reports of increasing insecurity to the personal data of users in the public and commercial domain applications has prompted the development of more robust and sound measures to protect the personal data of users from being stolen and spoofing. The present study aimed to introduce the scheme for integrating direct and indirect biometric key generation schemes with the application of Shamir‘s secret sharing algorithm in order to address the two disadvantages: revocability of the biometric key and the exception handling of biometric modality. This study used two different approaches for key generation using Shamir‘s secret sharing scheme: template based approach for indirect key generation and template-free. The findings of this study demonstrated that the encryption key generated by the proposed system was not required to be stored in the database which prevented the attack on the privacy of the data of the individuals from the hackers. Interestingly, the proposed system was also able to generate multiple encryption keys with varying lengths. Furthermore, the results of this study also offered the flexibility of providing the multiple keys for different applications for each user. The results from this study, consequently, showed the considerable potential and prospect of the proposed scheme to generate encryption keys directly and indirectly from the biometric samples, which could enhance its success in biometric security field.

This research dealt with the problem of neuronal membrane detection, in which the core challenge is distinguishing membranes from organelles. A simple and efficient optimisation framework is proposed based on several basic processing steps, including local contrast enhancement, denoising, thresholding, hole-filling, watershed segmentation, and morphological operations. The two main algorithms proposed Image Processing Chain Optimisation (IPCO) and Multiple IPCO (MIPCO)combine elements of Genetic Algorithms, Differential Evolution, and Rank-based uniform crossover. 91.67% is the highest recorded individual IPCO score with a speed of 280 s, and 92.11% is the highest recorded ensembles IPCO score whereas 91.80% is the highest recorded individual MIPCO score with a speed of 540 s for typically less than 500 optimisation generations and 92.63% is the highest recorded ensembles MIPCO score. Further, IPCO chains and MIPCO networks do not require specialised hardware and they are easy to use and deploy. This is the first application of this approach in the context of the Drosophila first instar larva ventral nerve cord. Both algorithms use existing image processing functions, but optimise the way in which they are configured and combined. The approach differs from related work in terms of the set of functions used, the parameterisations allowed, the optimisation methods adopted, the combination framework, and the testing and analyses conducted. Both IPCO and MIPCO are efficient and interpretable, and facilitate the generation of new insights. Systematic analyses of the statistics of optimised chains were conducted using 30 microscopy slices with corresponding ground truth. This process revealed several interesting and unconventional insights pertaining to preprocessing, classification, post-processing, and speed, and the appearance of functions in unorthodox positions in image processing chains, suggesting new sets of pipelines for image processing. One such insight revealed that, at least in the context of our membrane detection data, it is typically better to enhance, and even classify, data before denoising them.

La reconnaissance des biomètries douces (genre, âge, etc.)trouve ses applications dans plusieurs domaines. Les approches proposéesse basent sur l’analyse de l’apparence (images 2D), très sensiblesaux changements de la pose et à l’illumination, et surtout pauvre en descriptionsmorphologiques. Dans cette thèse, nous proposons d’exploiterla forme 3D du visage. Basée sur une approche Riemannienne d’analysede formes 3D, nous introduisons quatre descriptions denses à savoir: lasymétrie bilatérale, la moyenneté, la configuration spatiale et les variationslocales de sa forme. Les évaluations faites sur la base FRGCv2 montrentque l’approche proposée est capable de reconnaître des biomètries douces.A notre connaissance, c’est la première étude menée sur l’estimation del’âge, et c’est aussi la première étude qui propose d’explorer les corrélationsentre les attributs faciaux, à partir de formes 3D. / Soft-Biometric (gender, age, etc.) recognition has shown growingapplications in different domains. Previous 2D face based studies aresensitive to illumination and pose changes, and insufficient to representthe facial morphology. To overcome these problems, this thesis employsthe 3D face in Soft-Biometric recognition. Based on a Riemannian shapeanalysis of facial radial curves, four types of Dense Scalar Field (DSF) featuresare proposed, which represent the Averageness, the Symmetry, theglobal Spatiality and the local Gradient of 3D face. Experiments with RandomForest on the 3D FRGCv2 dataset demonstrate the effectiveness ofthe proposed features in Soft-Biometric recognition. Furtherly, we demonstratethe correlations of Soft-Biometrics are useful in the recognition. Tothe best of our knowledge, this is the first work which studies age estimation,and the correlations of Soft-Biometrics, using 3D face.

Feature recognition aids CAD model simplification in engineering analysis and machining path in manufacturing. In the domain of CAD model simplification, classic feature recognition approaches face two challenges: 1) insufficient performances; 2) engineering features are diverse, and no system can hard-code all possible features in advance. A declarative approach allows engineers to specify new features without having to design algorithms to find them. However, naive translation of declarations leads to executable algorithms with high time complexity. Inspired by relational database management systems (RDBMS), I suppose that if there exists a way to turn a feature declaration into an SQL query that is further processed by a database engine interfaced to a CAD modeler, the optimizations can be utilized for “free”. Testbeds are built to verify the idea. Initially, I devised a straightforward translator to turn feature declarations into queries. Experiments on SQLite show it gives a quasiquadratic performance for common features. Then it is extended with a new translator and PostgreSQL. In the updated version, I have made a significant breakthrough – my approach is the first to achieve linear time performance with respect to model size for common features, and acceptable times for real industrial models. I learn from the testbeds that PostgreSQL uses hash joins reduce the search space enable a fast feature finding. Besides, I have further improved the performance by: (i) lazy evaluation, which can be used to reduce the workload on the CAD modeler, and (ii) predicate ordering, which reorders the query plan by taking into account the time needed to compute various geometric operations. Experimental results are presented to validate their benefits.

Dans un contexte économique où les concurrences internationales sont exacerbées, la customisation, ou personnalisation de masse des produits devient aujourd’hui une stratégie très importante des entreprises pour améliorer la valeur perçue de leurs produits. Cependant, les expériences des plateformes de customisations actuelles en ligne ne sont pas pleinement satisfaisantes car les choix personnalisés sont essentiellement limitées à des couleurs et à des motifs. Les dimensions sensorielles des produits, incluant en particulier l’apparence et le toucher des matières tout autant que le bien-aller du vêtement sont rarement proposés.Dans le cadre de ma thèse doctorale, nous avons proposé une plateforme de co-création, permettant aux commerçants, aux créateurs et aux clients d’acquérir conjointement une nouvelle expérience sur le développement de vêtements personnalisés à la valeur ajoutée plus élevée sans entraîner de surcoûts industriels. La construction de cette plateforme consiste en plusieurs parties. Tout à bord, nous avons sélectionné, par une expérience sensorielle, un logiciel de CAO en confection 3D bien adapté en termes de la qualité de rendu du vêtement virtuel. Ensuite, nous avons proposé un plan d’expérience sensorielle par utilisation d’une nouvelle méthode d’apprentissage actif proposée afin d’acquérir, sans mesures physiques, les paramètres techniques de l’étoffe dans un délai très court. Cette méthode est efficace, rapide, facile à réaliser et notamment très significative pour des transactions des textiles en ligne. Puis nous avons caractérisé quantitativement la perception du vêtement virtuel par des notes numériques sur un ensemble de descripteurs sensoriels normalisés. Ces derniers concernent l’apparence et le toucher de la matière, ainsi que le fit du vêtement. Les données sensorielles ont été obtenues respectivement dans deux autres expériences sensorielles. Par apprentissage de ces données, nous avons établi deux modèles. Le premier permet de caractériser la relation entre la perception sur l’apparence et le toucher du matériau virtuel et les paramètres techniques correspondants, constituant une entrée du logiciel de CAO en confection. Le deuxième modèle permet de caractériser la relation entre la perception du fit du vêtement virtuel et les paramètres des patrons. A l'aide des deux modèles précédents, les créateurs et consommateurs peuvent ajuster les éléments initiaux de conception pour les matières et les patrons du vêtement selon leurs attentes au niveau du sensoriel. / Under the exacerbated worldwide competition, the mass customization or personalization of products is now becoming an important strategy for companies to enhance the perceived value of their products. However, the current online customization experiences are not fully satisfying for consumers because the choices are mostly limited to colors and motifs. The sensory fields of products, particularly the material’s appearance and hand as well as the garment fit are barely concerned.In my PhD research project, we have proposed a new collaborative design platform. It permits merchants, designers and consumers to have a new experience during the development of highly valued personalized garments without extra industrial costs. The construction of this platform consists of several parts. At first, we have selected, through a sensory experiment, an appropriate 3D garment CAD software in terms of rending quality. Then we have proposed an active leaning-based experimental design in order to find the most appropriate values of the fabric technical parameters permitting to minimize the overall perceptual difference between real and virtual fabrics in static and dynamic scenarios. Afterwards, we have quantitatively characterized the human perception on virtual garment by using a number of normalized sensory descriptors. These descriptors involve not only the appearance and the hand of the fabric but also the garment fit. The corresponding sensory data have been collected through two sensory experiments respectively. By learning from the experimental data, two models have been established. The first model permits to characterize the relationship between the appearance and hand perception of virtual fabrics and corresponding technical parameters that constitute the inputs of the 3D garment CAD software. The second model concerns the relationship between virtual garment fit perception and the pattern design parameters. These two models constitute the main components of the collaborative design platform. Using this platform, we have realized a number of garments meeting consumer’s personalized perceptual requirements.

An investigation has been made into symmetry features of patterns as a means by which the patterns are described, or with which they are transformed prior to classification in order to assist a pattern recognition system. There are two main points of departure from existing symmetry use in the pattern recognition domain. The first is the adoption of the theory that patterns can be categorised solely using a map of the symmetry features that exist within the static pattern. The second is the application of symmetry transforms to aid non-trivial recognition in patterns which are not intended to be perfectly symmetrical. An experiment is conducted to classify the reflectional symmetry features of digits, using the Generalised Symmetry Transform to produce the features and Probabilistic Neural Networks to perform the classification. Symmetry feature information is also used to define parameters of affine transformations to achieve improved performance in tolerance to variances in position and orientation. The results lead to an investigation of the role of asymmetry. The Generalised Symmetry Transform is modified to produce two related transforms: the Generalised Asymmetry Transform and the Generalised Asymmetry and Symmetry Transform. Finally, a new symmetry transform is proposed which separates the factors affecting the degree of symmetry in an image into three non-linear functions of corresponding pairs of pixels. These factors are: the colour intensity values; the pixel orientations; and the respective distance between the point and potential reflection plane. The strictness of symmetry, or tolerance to non-symmetrical artifacts, is defined in variable parameters which are set to suit the desired application. This new transform is called the Reflectional Symmetry Transform. The structure of its input and output match those of the Generalised Symmetry Transform, which it is intended to replace.

Recognising humans by the way they walk has attracted a significant interest in recent years due to its potential use in a number of applications such as automated visual surveillance. Technologies utilising gait biometrics have the potential to provide safer society and improve quality of life. However, automated gait recognition is a very challenging research problem and some fundamental issues remain unsolved. At the moment, gait recognition performs well only when samples acquired in similar conditions are matched. An operational automated gait recognition system does not yet exist. The primary aim of the research presented in this thesis is to understand the main challenges associated with deployment of gait recognition and to propose novel solutions to some of the most fundamental issues. There has been lack of understanding of the effect of some subject dependentcovariates on gait recognition performance. We have proposed a novel dataset that allows analyses of various covariates in a principled manner. The results of thedatabase evaluation revealed that elapsed time does not affect recognition in the short to medium term, contrary to what other studies have concluded. The analyses show how other factors related to the subject affect recognition performance. Only few gait recognition approaches have been validated in real world conditions. We have collected a new dataset at two realistic locations. Using the database we have shown that there are many environment related factors that can affect performance. The quality of silhouettes has been identified as one of the most important issues for translating gait recognition research to the ‘real-world’. The existing quality algorithms proved insufficient and therefore we extended quality metrics and proposed new ways of improving signature quality and therefore performance. A new fully working automated system has been implemented. Experiments using the system in ‘real-world’ conditions have revealed additional challenges not present when analysing datasets of fixed size. In conclusion, the research has investigated many of the factors that affect current gait recognition algorithms and has presented novel approaches of dealing with some of the most important issues related to translating gait recognition to real-world environments.