|
Subspace Modeling of Prosodic Features for Speaker Verification
Kockmann, Marcel ; Kenny, Patrick (oponent) ; Nöth, Elmar (oponent) ; Černocký, Jan (vedoucí práce)
The thesis investigates into speaker verification by means of prosodic features. This includes an appropriate representation of speech by measurements of pitch, energy and duration of speech sounds. Two diverse parameterization methods are investigated: the first leads to a low-dimensional well-defined set, the second to a large-scale set of heterogeneous prosodic features. The first part of this work concentrates on the development of so called prosodic contour features. Different modeling techniques are developed and investigated, with a special focus on subspace modeling. The second part focuses on a novel subspace modeling technique for the heterogeneous large-scale prosodic features. The model is theoretically derived and experimentally evaluated on official NIST Speaker Recognition Evaluation tasks. Huge improvements over the current state-of-the-art in prosodic speaker verification were obtained. Eventually, a novel fusion method is presented to elegantly combine the two diverse prosodic systems. This technique can also be used to fuse the higher-level systems with a high-performing cepstral system, leading to further significant improvements.
|
|
Acceleration of Object Detection Using Classifiers
Juránek, Roman ; Kälviäinen, Heikki (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
Detection of objects in computer vision is a complex task. One of most popular and well explored approaches is the use of statistical classifiers and scanning windows. In this approach, classifiers learned by AdaBoost algorithm (or some modification) are often used as they achieve low error rates, high detection rates and they are suitable for detection in real-time applications. Object detection run-time which uses such classifiers can be implemented by various methods and properties of underlying architecture can be used for speed-up of the detection. For the purpose of acceleration, graphics hardware, multi-core architectures, SIMD or other means can be used. The detection is often implemented on programmable hardware. The contribution of this thesis is to introduce an optimization technique which enhances object detection performance with respect to an user defined cost function. The optimization balances computations of previously learned classifiers between two or more run-time implementations in order to minimize the cost function. The optimization method is verified on a basic example -- division of a classifier to a pre-processing unit implemented in FPGA, and a post-processing unit in standard PC.
|
|
Optimization of Gaussian Mixture Subspace Models and Related Scoring Algorithms in Speaker Verification
Glembek, Ondřej ; Brummer, Niko (oponent) ; Campbell,, William (oponent) ; Burget, Lukáš (vedoucí práce)
This thesis deals with Gaussian Mixture Subspace Modeling in automatic speaker recognition. The thesis consists of three parts. In the first part, Joint Factor Analysis (JFA) scoring methods are studied. The methods differ mainly in how they deal with the channel of the tested utterance. The general JFA likelihood function is investigated and the methods are compared both in terms of accuracy and speed. It was found that linear approximation of the log-likelihood function gives comparable results to the full log-likelihood evaluation while simplyfing the formula and dramatically reducing the computation speed. In the second part, i-vector extraction is studied and two simplification methods are proposed. The motivation for this part was to allow for using the state-of-the-art technique on small scale devices and to setup a simple discriminative-training system. It is shown that, for long utterances, while sacrificing the accuracy, we can get very fast and compact i-vector systems. On a short-utterance(5-second) task, the results of the simplified systems are comparable to the full i-vector extraction. The third part deals with discriminative training in automatic speaker recognition. Previous work in the field is summarized and---based on the knowledge from the earlier chapters of this work---discriminative training of the i-vector extractor parameters is proposed. It is shown that discriminative re-training of the i-vector extractor can improve the system if the initial estimation is computed using the generative approach.
|
|
Analysis and Testing of Concurrent Programs
Letko, Zdeněk ; Lourenco, Joao (oponent) ; Sekanina, Lukáš (oponent) ; Vojnar, Tomáš (vedoucí práce)
The thesis starts by providing a taxonomy of concurrency-related errors and an overview of their dynamic detection. Then, concurrency coverage metrics which measure how well the synchronisation and concurrency-related behaviour of tested programs has been examined are proposed together with a~methodology for deriving such metrics. The proposed metrics are especially suitable for saturation-based and search-based testing. Next, a novel coverage-based noise injection techniques that maximise the number of interleavings witnessed during testing are proposed. A comparison of various existing noise injection heuristics and the newly proposed heuristics on a set of benchmarks is provided, showing that the proposed techniques win over the existing ones in some cases. Finally, a novel use of stochastic optimisation algorithms in the area of concurrency testing is proposed in the form of their application for finding suitable combinations of values of the many parameters of tests and the noise injection techniques. The approach has been implemented in a prototype way and tested on a set of benchmark programs, showing its potential to significantly improve the testing process.
|
|
Relational Verification of Programs with Integer Data
Konečný, Filip ; Bouajjani, Ahmed (oponent) ; Jančar, Petr (oponent) ; Vojnar, Tomáš (vedoucí práce)
This work presents novel methods for verification of reachability and termination properties of programs that manipulate unbounded integer data. Most of these methods are based on acceleration techniques which compute transitive closures of program loops. We first present an algorithm that accelerates several classes of integer relations and show that the new method performs up to four orders of magnitude better than the previous ones. On the theoretical side, our framework provides a common solution to the acceleration problem by proving that the considered classes of relations are periodic. Subsequently, we introduce a semi-algorithmic reachability analysis technique that tracks relations between variables of integer programs and applies the proposed acceleration algorithm to compute summaries of procedures in a modular way. Next, we present an alternative approach to reachability analysis that integrates predicate abstraction with our acceleration techniques to increase the likelihood of convergence of the algorithm. We evaluate these algorithms and show that they can handle a number of complex integer programs where previous approaches failed. Finally, we study the termination problem for several classes of program loops and show that it is decidable. Moreover, for some of these classes, we design a polynomial time algorithm that computes the exact set of program configurations from which nonterminating runs exist. We further integrate this algorithm into a semi-algorithmic method that analyzes termination of integer programs, and show that the resulting technique can verify termination properties of several non-trivial integer programs.
|
|
Acceleration Methods for Evolutionary Design of Digital Circuits
Vašíček, Zdeněk ; Miller, Julian (oponent) ; Zelinka,, Ivan (oponent) ; Sekanina, Lukáš (vedoucí práce)
Although many examples showing the merits of evolutionary design over conventional design techniques utilized in the field of digital circuits design have been published, the evolutionary approaches are usually hardly applicable in practice due to the various so-called scalability problems. The scalability problem represents a general problem that refers to a situation in which the evolutionary algorithm is able to provide a solution to a small problem instances only. For example, the scalability of evaluation of a candidate digital circuit represents a serious issue because the time needed to evaluate a candidate solution grows exponentially with the increasing number of primary inputs. In this thesis, the scalability problem of evaluation of a candidate digital circuit is addressed. Three different approaches to overcoming this problem are proposed. Our goal is to demonstrate that the evolutionary design approach can produce interesting and human competitive solutions when the problem of scalability is reduced and thus a sufficient number of generations can be utilized. In order to increase the performance of the evolutionary design of image filters, a domain specific FPGA-based accelerator has been designed. The evolutionary design of image filters is a kind of regression problem which requires to evaluate a large number of training vectors as well as generations in order to find a satisfactory solution. By means of the proposed FPGA accelerator, very efficient nonlinear image filters have been discovered. One of the discovered implementations of an impulse noise filter consisting of four evolutionary designed filters is protected by the Czech utility model. A different approach has been introduced in the area of logic synthesis. A method combining formal verification techniques with evolutionary design that allows a significant acceleration of the fitness evaluation procedure was proposed. The proposed system can produce complex and simultaneously innovative designs, overcoming thus the major bottleneck of the evolutionary synthesis at gate level. The proposed method has been evaluated using a set of benchmark circuits and compared with conventional academia as well as commercial synthesis tools. In comparison with the conventional synthesis tools, the average improvement in terms of the number of gates provided by our system is approximately 25%. Finally, the problem of the multiple constant multiplier design, which belongs to the class of problems where a candidate solution can be perfectly evaluated in a short time, has been investigated. We have demonstrated that there exists a class of circuits that can be evaluated efficiently if a domain knowledge is utilized (in this case the linearity of components).
|
|
Simulace a protiřetězce pro efektivní práci s konečnými automaty
Holík, Lukáš ; Černá, Ivana (oponent) ; Jančar, Petr (oponent) ; Vojnar, Tomáš (vedoucí práce)
This thesis is focused on techniques for finite automata and their use in practice, with the main emphasis on nondeterministic tree automata. This concerns namely techniques for size reduction and language inclusion testing, which are two problems that are crucial for many applications of tree automata. For size reduction of tree automata, we adapt the simulation quotient technique that is well established for finite word automata. We give efficient algorithms for computing tree automata simulations and we also introduce a new type of relation that arises from a combination of tree automata downward and upward simulation and that is very well suited for quotienting. The combination principle is relevant also for word automata. We then generalise the so called antichain universality and language inclusion checking technique developed originally for finite word automata for tree automata. Subsequently, we improve the antichain technique for both word and tree automata by combining it with the simulation-based inclusion checking techniques, significantly improving efficiency of the antichain method. We then show how the developed reduction and inclusion checking methods improve the method of abstract regular tree model checking, the method that was the original motivation for starting the work on tree automata. Both the reduction and the language inclusion methods are based on relatively simple and general principles that can be further extended for other types of automata and related formalisms. An example is our adaptation of the reduction methods for alternating Büchi automata, which results in an efficient alternating automata size reduction technique.
|
|
On-line Data Analysis Based on Visual Codebooks
Beran, Vítězslav ; Honec, Jozef (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
This work introduces the new adaptable method for on-line video searching in real-time based on visual codebook. The new method addresses the high computational efficiency and retrieval performance when used on on-line data. The method originates in procedures utilized by static visual codebook techniques. These standard procedures are modified to be able to adapt to changing data. The procedures, that improve the new method adaptability, are dynamic inverse document frequency, adaptable visual codebook and flowing inverted index. The developed adaptable method was evaluated and the presented results show how the adaptable method outperforms the static approaches when evaluating on the video searching tasks. The new adaptable method is based on introduced flowing window concept that defines the ways of selection of data, both for system adaptation and for processing. Together with the concept, the mathematical background is defined to find the best configuration when applying the concept to some new method. The practical application of the adaptable method is particularly in the video processing systems where significant changes of the data domain, unknown in advance, is expected. The method is applicable in embedded systems monitoring and analyzing the broadcasted TV on-line signals in real-time.
|
|
HUMAN ACTION RECOGNITION IN VIDEO
Řezníček, Ivo ; Baláž, Teodor (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
This thesis focuses on the improvement of human action recognition systems. It reviews the state-of-the-art in the field of action recognition from video. It describes techniques of digital image and video capture, and explains computer representations of image and video. This thesis further describes how local feature vectors and local space-time feature vectors are used, and how captured data is prepared for further analysis, such as classification methods. This is typically done with video segments of arbitrarily varying length. The key contribution of this work explores the hypothesis that the analysis of different types of actions requires different segment lenghts to achieve optimal quality of recognition. An algorithm to find these optimal lengths is proposed, implemented, and tested. Using this algorithm, the hypothesis was experimentally proven. It was also shown that by finding the optimal length, the prediction and classification power of current algorithms is improved upon. Supporting experiments, results, and proposed exploitations of these findings are presented.
|
|
Multimedia Data Processing in Heterogeneous Distributed Environment
Kajan, Rudolf ; Ferko,, Andrej (oponent) ; Míkovec, Zdeněk (oponent) ; Herout, Adam (vedoucí práce)
Ubiquitous computing, a paradigm in which the processing of information is linked with each activity or object as encountered, was proposed by Mark Weiser as the next era for interacting with computers. Its goal is to enable people to interact with devices more naturally and casually in ways that suit whatever location or context they find themselves in. Ubiquitous computing focuses on learning by removing the complexity of computing and increases efficiency while using computing for different daily activities. But after more than 15 years since Weiser formulated these goals, several aspects of ubiquitous computing are still not a part of user experience with today’s technology. Seamless integration with environment leading to technological invisibility or user interaction spanning across multiple devices pose still a great challenge. The main goal of our work is to make a step towards making the idea of ubiquitous computing a reality by addressing the question about intuitive information sharing between a personal device and a situated display. We have developed three interaction techniques which support unobtrusive content exchange between touch-enabled personal device and a large display - whether it is shared-private or public. These techniques are based on video streams, augmented reality, and analysis of gaze data. Besides the interaction techniques, we also present a framework for real-time application state acquisition and reconstruction on target platform. We report on user studies focused on the usability of our prototypes and a system performance evaluations. Our experiments were formed around real-life scenarios which are commonly experienced throughout the day. For interactions based on video streams, the results indicate that our techniques outperform the existing solutions: the localization and task migration is done in real time on a midlevel cellphone; the localization is reliable even for different observation angles and for cluttered screen content. Our technique based on gaze analysis goes even further by allowing for modeling of implicit user preferences through gaze data, while being accurate and unobtrusive.
|
host ::
