In this thesis, a general solution to the component separation problem in images is introduced. Unlike most existing works, the spatial dependencies of images are modelled in the separation process with the use of Markov random fields (MRFs). In the MRFs model, Cauchy density is used for the gradient images. We provide a general Bayesian framework for the estimation of the parameters of this model. Due to the intractability of the problem we resort to numerical solutions for the joint maximization of the a posteriori distribution of the sources, the mixing matrix and the noise variances. For numerical solution, four different methods are proposed. In first method, the difficulty of working analytically with general Gibbs distributions of MRF is overcome by using an approximate density. In this approach, the Gibbs distribution is modelled by the product of directional Gaussians. The ...

This Thesis is focused on the combination of multiple biometric traits for automatic person authentication, in what is called a multimodal biometric system. More generally, any type of biometric information can be combined in what is called a multibiometric system. The information sources in multibiometrics include not only multiple biometric traits but also multiple sensors, multiple biometric instances (e.g., different fingers in fingerprint verification), repeated instances, and multiple algorithms. Most of the approaches found in the literature for combining these various information sources are based on the combination of the matching scores provided by individual systems built on the different biometric evidences. The combination schemes following this architecture are typically based on combination rules or trained pattern classifiers, and most of them assume that the score level fusion function is fixed at verification time. This Thesis considers the problem of ...

The unprecedented processing demand, posed by the explosion of big data, challenges researchers to design efficient and adaptive machine learning algorithms that do not require persistent retraining and avoid learning redundant information. Inspired from learning techniques of intelligent biological agents, identifying transferable knowledge across learning problems has been a significant research focus to improve machine learning algorithms. In this thesis, we address the challenges of knowledge transfer through embedding spaces that capture and store hierarchical knowledge. In the first part of the thesis, we focus on the problem of cross-domain knowledge transfer. We first address zero-shot image classification, where the goal is to identify images from unseen classes using semantic descriptions of these classes. We train two coupled dictionaries which align visual and semantic domains via an intermediate embedding space. We then extend this idea by training deep networks that ...

Detecting anomalous sounds is a difficult task: First, audio data is very high-dimensional and anomalous signal components are relatively subtle in relation to the entire acoustic scene. Furthermore, normal and anomalous audio signals are not inherently different because defining these terms strongly depends on the application. Third, usually only normal data is available for training a system because anomalies are rare, diverse, costly to produce and in many cases unknown in advance. Such a setting is called semi-supervised anomaly detection. In domain-shifted conditions or when only very limited training data is available, all of these problems are even more severe. The goal of this thesis is to overcome these difficulties by teaching an embedding model to learn data representations suitable for semi-supervised anomalous sound detection. More specifically, an anomalous sound detection system is designed such that the resulting representations of ...

The objectives of the research were to develop embedded and non-embedded lossy coding algorithms for images based on lattice vector quantisation and the discrete wavelet transform. We also wanted to develop context-based entropy coding methods (as opposed to simple first order entropy coding). The main objectives can therefore be summarised as follows: (1) To develop algorithms for intra and inter-band formed vectors (vectors with coefficients from the same sub-band or across different sub-bands) which compare favourably with current high performance wavelet based coders both in terms of rate/distortion performance of the decoded image and also subjective quality; (2) To develop new context-based coding methods (based on vector quantisation). The alternative algorithms we have developed fall into two categories: (a) Entropy coded and Binary uncoded successive approximation lattice vector quantisation (SALVQ- E and SA-LVQ-B) based on quantising vectors formed intra-band. This ...

In the realm of traffic monitoring, efficient video analysis is paramount yet challenging due to intensive computational demands. This thesis addresses this issue by introducing novel methods to operate in the compressed domain. Four methods are proposed for image reconstruction from High Efficiency Video Coding (HEVC) Intra bitstreams, namely, the Block Partition Based Method (Mbp), the Prediction Unit Based Method (Mpu), the Random Perturbation Based Method (Mrp), and the Luma based method (My). These methods aim to provide a compact representation of the original image while retaining relevant information for video understanding tasks. Our methods substantially reduce data transmission requirements and memory footprint. Specifically, images created via Mbp and Mpu require 1/1,536 and 1/192 of the memory needed by pixel domain images, respectively. Moreover, these methods offer computational speedup between 1.25 to 4 times, yielding efficiencies in video analysis. The ...

Every day we share our personal information through digital systems which are constantly exposed to threats. For this reason, security-oriented disciplines of signal processing have received increasing attention in the last decades: multimedia forensics, digital watermarking, biometrics, network monitoring, steganography and steganalysis are just a few examples. Even though each of these elds has its own peculiarities, they all have to deal with a common problem: the presence of one or more adversaries aiming at making the system fail. Adversarial Signal Processing lays the basis of a general theory that takes into account the impact that the presence of an adversary has on the design of effective signal processing tools. By focusing on the application side of Adversarial Signal Processing, namely adversarial information fusion in distributed sensor networks, and adopting a game-theoretic approach, this thesis contributes to the above mission ...

Earth observation and land-cover analysis became a reality in the last 2-3 decades thanks to NASA airborne and spacecrafts such as Landsat. Inclusion of Hyperspectral Imaging (HSI) technology in some of these platforms has made possible acquiring large data sets, with high potential in analytical tasks but at the cost of advanced signal processing. In this thesis, effective/efficient feature extraction methods are proposed. Initially, contributions are introduced for efficient computation of the covariance matrix widely used in data reduction methods such as Principal Component Analysis (PCA). By taking advantage of the cube structure in HSI, onsite and real-time covariance computation is achieved, reducing memory requirements as well. Furthermore, following the PCA algorithm, a novel method called Folded-PCA (Fd-PCA) is proposed for efficiency while extracting both global and local features within the spectral pixels, achieved by folding the spectral samples from ...

Air traffic control (ATC) voice radio communication between aircraft pilots and controllers is subject to technical and functional constraints owing to the legacy radio system currently in use worldwide. This thesis investigates the embedding of digital side information, so called watermarks, into speech signals. Applied to the ATC voice radio, a watermarking system could overcome existing limitations, and ultimately increase safety, security and efficiency in ATC. In contrast to conventional watermarking methods, this field of application allows embedding of the data in perceptually irrelevant signal components. We show that the resulting theoretical watermark capacity far exceeds the capacity of conventional watermarking channels. Based on this finding, we present a general purpose blind speech watermarking algorithm that embeds watermark data in the phase of non-voiced speech segments by replacing the excitation signal of an autoregressive signal representation. Our implementation embeds the ...

During the past two decades, image compression has developed from a mostly academic Rate-Distortion (R-D) field, into a highly commercial business. Various lossless and lossy image coding techniques have been developed. This thesis represents an interdisciplinary work between the field of astronomy and digital image processing and brings new aspects into both of the fields. In fact, image compression had its beginning in an American space program for efficient data storage. The goal of this research work is to recognize and develop new methods for space observatories and software tools to incorporate compression in space astronomy standards. While the astronomers benefit from new objective processing and analysis methods and improved efficiency and quality, for technicians a new field of application and research is opened. For validation of the processing results, the case of InfraRed (IR) astronomy has been specifically analyzed. ...

This study proposes a new spectral representation called the Zeros of Z-Transform (ZZT), which is an all-zero representation of the z-transform of the signal. In addition, new chirp group delay processing techniques are developed for analysis of resonances of a signal. The combination of the ZZT representation with the chirp group delay processing algorithms provides a useful domain to study resonance characteristics of source and filter components of speech. Using the two representations, effective algorithms are developed for: source-tract decomposition of speech, glottal flow parameter estimation, formant tracking and feature extraction for speech recognition. The ZZT representation is mainly important for theoretical studies. Studying the ZZT of a signal is essential to be able to develop effective chirp group delay processing methods. Therefore, first the ZZT representation of the source-filter model of speech is studied for providing a theoretical background. ...

Nowadays, Nuclear Magnetic Resonance (NMR) is widely used in oncology as a non-invasive diagnostic tool in order to detect the presence of tumor regions in the human body. An application of NMR is Magnetic Resonance Imaging, which is applied in routine clinical practice to localize tumors and determine their size. Magnetic Resonance Imaging is able to provide an initial diagnosis, but its ability to delineate anatomical and pathological information is significantly improved by its combination with another NMR application, namely Magnetic Resonance Spectroscopy. The latter reveals information on the biochemical profile tissues, thereby allowing clinicians and radiologists to identify in a non{invasive way the different tissue types characterizing the sample under investigation, and to study the biochemical changes underlying a pathological situation. In particular, an NMR application exists which provides spatial as well as biochemical information. This application is called ...

Learning parameters from voluminous data can be prohibitive in terms of memory and computational requirements. Furthermore, new challenges arise from modern database architectures, such as the requirements for learning methods to be amenable to streaming, parallel and distributed computing. In this context, an increasingly popular approach is to first compress the database into a representation called a linear sketch, that satisfies all the mentioned requirements, then learn the desired information using only this sketch, which can be significantly faster than using the full data if the sketch is small. In this thesis, we introduce a generic methodology to fit a mixture of probability distributions on the data, using only a sketch of the database. The sketch is defined by combining two notions from the reproducing kernel literature, namely kernel mean embedding and Random Features expansions. It is seen to correspond ...

Single-pixel imaging is a recent paradigm that allows the acquisition of images at reasonably low cost by exploiting hardware compression of the data. The architecture of a single-pixel camera consists of only two elements: a spatial light modulator, and a single-point detector. The key idea is to measure the projection at the detector (i.e., the inner product) of the scene under view -the image- with some patterns. The post-processing of a sequence of measurements obtained with different patterns permits the restoring of the desired image. Single-pixel imaging has several advantages, which are of interest for different applications, and especially in the biomedical field. In particular, a time-resolved single-pixel imaging system benefits fluorescence lifetime sensing. Such a set-up can be coupled to a spectrometer, to supplement the lifetime with spectral information. However, the main limitation of single-pixel imaging is the speed ...

This thesis is concerned with the design of a complete framework that allows the real-time recognition of humans in a video stream acquired by a static camera. For each stage of the processing chain, which takes as input the raw images of the stream and eventually outputs the identity of the persons, we propose an original algorithm. The first algorithm is a background subtraction technique named ViBe. The purpose of ViBe is to detect the parts of the images that contain moving objects. The second algorithm determines which moving objects correspond to individuals. The third algorithm allows the recognition of the detected individuals from their gait. Our background subtraction algorithm, ViBe, uses a collection of samples to model the history of each pixel. The current value of a pixel is classified by comparison with the closest samples that belong to ...