This work is about suggestion of the software for neural network based image segmentation. It defines basic terms for this topics. It is focusing mainly at preperation imaging information for image segmentation using neural network. It describes and compares different aproaches for image segmentation.

<p>Energy minimization approach is widely used in image processing applications.<br />Many image processing problems can be modelled in a form of a minimization<br />problem. This thesis deals with two crucial tasks of image analysis workflows:<br />image restoration and segmentation of images corrupted by blur and noise. Both<br />image restoration and segmentation are modelled as energy minimization<br />problems, where energy function is composed of two parts: data fidelity term and<br />regularization term. The main contribution of this thesis is development of new<br />data fidelity and regularization terms for both image restoration and<br />segmentation tasks.<br />Image restoration methods (non-blind and blind deconvolution and superresolution<br />reconstruction) developed within this thesis are suited for mixed<br />Poisson-Gaussian noise which is encountered in many realistic imaging<br />conditions. We use generalized Anscombe variance stabilization transformation<br />for removing signal-dependency of noise. We propose novel data fidelity term<br />which incorporates variance stabilization transformation process into account.<br />Turning our attention to the regularization term for image restoration, we<br />investigate how sparsity promoting regularization in the gradient domain<br />formulated as Total Variation, can be improved in the presence of blur and mixed<br />Poisson-Gaussian noise. We found that Huber potential function leads to<br />significant improvement of restoration performance.<br />In this thesis we propose new segmentation method, the so called coverage<br />segmentation, which estimates the relative coverage of each pixel in a sensed<br />image by each image component. Its data fidelity term takes into account<br />blurring and down-sampling processes and in that way it provides robust<br />segmentation in the presence of blur, allowing at the same time segmentation at<br />increased spatial resolution. In addition, new sparsity promoting regularization<br />terms are suggested: (i) Huberized Total Variation which provides smooth object<br />boundaries and noise removal, and (ii) non-edge image fuzziness, which<br />responds to an assumption that imaged objects are crisp and that fuzziness is<br />mainly due to the imaging and digitization process.<br />The applicability of here proposed restoration and coverage segmentation<br />methods is demonstrated for Transmission Electron Microscopy image<br />enhancement and segmentation of micro-computed tomography and<br />hyperspectral images.</p> / <p>Поступак минимизације функције енергије је често коришћен за<br />решавање проблема у обради дигиталне слике. Предмет истраживања<br />тезе су два круцијална задатка дигиталне обраде слике: рестаурација и<br />сегментација слика деградираних шумом и замагљењем. И рестaурација<br />и сегментација су моделовани као проблеми минимизације функције<br />енергије која представља збир две функције: функције фитовања<br />података и регуларизационе функције. Главни допринос тезе је развој<br />нових функција фитовања података и нових регуларизационих функција<br />за рестаурацију и сегментацију.<br />Методе за рестаурацију (оне код којих је функција замагљења позната и<br />код којих је функцију замагљења потребно оценити на основу датих<br />података као и методе за реконструкцију слике у супер-резолуцији)<br />развијене у оквиру ове тезе третирају мешавину Поасоновог и Гаусовог<br />шума који се појављује у многобројним реалистичним сценаријима. За<br />третирање такве врсте шума користили смо нелинеарну трансформацију<br />и предложили смо нову функцију фитовања података која узима у обзир<br />такву трансформацију. У вези са регуларизационим функцијама смо<br />тестирали хипотезу да се функција Тоталне Варијације која промовише<br />ретку слику у градијентном домену може побољшати уколико се користе<br />тзв. потенцијалне функције. Показали смо да се употребом Хуберове<br />потенцијалне функције може значајно побољшати квалитет рестауриране<br />слике која је деградирана замагљењем и мешавином Поасоновог и<br />Гаусовог шума.<br />У оквиру тезе смо предложили нову методу сегментације која допушта<br />делимичну покривеност пиксела објектом. Функција фитовања података<br />ове методе укључује и модел замагљења и смањења резолуције. На тај<br />начин је постигнута робустност сегментације у присуству замагљења и<br />добијена могућност сегментирања слике у супер-резолуцији. Додатно,<br />нове регуларизационе функције које промовишу ретке репрезентације<br />слике су предложене.<br />Предложене методе рестаурације и сегментације која допушта делимичну<br />покривеност пиксела објектом су примењене на слике добијене помоћу<br />електронског микроскопа, хиперспектралне слике и медицинске ЦТ слике.</p> / <p>Postupak minimizacije funkcije energije je često korišćen za<br />rešavanje problema u obradi digitalne slike. Predmet istraživanja<br />teze su dva krucijalna zadatka digitalne obrade slike: restauracija i<br />segmentacija slika degradiranih šumom i zamagljenjem. I restauracija<br />i segmentacija su modelovani kao problemi minimizacije funkcije<br />energije koja predstavlja zbir dve funkcije: funkcije fitovanja<br />podataka i regularizacione funkcije. Glavni doprinos teze je razvoj<br />novih funkcija fitovanja podataka i novih regularizacionih funkcija<br />za restauraciju i segmentaciju.<br />Metode za restauraciju (one kod kojih je funkcija zamagljenja poznata i<br />kod kojih je funkciju zamagljenja potrebno oceniti na osnovu datih<br />podataka kao i metode za rekonstrukciju slike u super-rezoluciji)<br />razvijene u okviru ove teze tretiraju mešavinu Poasonovog i Gausovog<br />šuma koji se pojavljuje u mnogobrojnim realističnim scenarijima. Za<br />tretiranje takve vrste šuma koristili smo nelinearnu transformaciju<br />i predložili smo novu funkciju fitovanja podataka koja uzima u obzir<br />takvu transformaciju. U vezi sa regularizacionim funkcijama smo<br />testirali hipotezu da se funkcija Totalne Varijacije koja promoviše<br />retku sliku u gradijentnom domenu može poboljšati ukoliko se koriste<br />tzv. potencijalne funkcije. Pokazali smo da se upotrebom Huberove<br />potencijalne funkcije može značajno poboljšati kvalitet restaurirane<br />slike koja je degradirana zamagljenjem i mešavinom Poasonovog i<br />Gausovog šuma.<br />U okviru teze smo predložili novu metodu segmentacije koja dopušta<br />delimičnu pokrivenost piksela objektom. Funkcija fitovanja podataka<br />ove metode uključuje i model zamagljenja i smanjenja rezolucije. Na taj<br />način je postignuta robustnost segmentacije u prisustvu zamagljenja i<br />dobijena mogućnost segmentiranja slike u super-rezoluciji. Dodatno,<br />nove regularizacione funkcije koje promovišu retke reprezentacije<br />slike su predložene.<br />Predložene metode restauracije i segmentacije koja dopušta delimičnu<br />pokrivenost piksela objektom su primenjene na slike dobijene pomoću<br />elektronskog mikroskopa, hiperspektralne slike i medicinske CT slike.</p>

Recently, Deep Learning has revolutionized many fields, where one such area is Voice Activity Detection (VAD). This is of great interest to sectors of society concerned with detecting speech in sound signals. One such sector is the police, where criminal investigations regularly involve analysis of audio material. Convolutional Neural Networks (CNN) have recently become the state-of-the-art method of detecting speech in audio. But so far, understanding the impact of noise and sampling rates on such methods remains incomplete. Additionally, there are evaluation metrics from neighboring fields that remain unintegrated into VAD. We trained on four different sampling rates and found that changing the sampling rate could have dramatic effects on the results. As such, we recommend explicitly evaluating CNN-based VAD systems on pertinent sampling rates. Further, with increasing amounts of white Gaussian noise, we observed better performance by increasing the capacity of our Gated Recurrent Unit (GRU). Finally, we discuss how careful consideration is necessary when choosing a main evaluation metric, leading us to recommend Polyphonic Sound Detection Score (PSDS).

This thesis aims to contribute to the developments of wireless communication systems. The work generally consists of three parts: the first part is a discussion on general digital communication systems, the second part focuses on wireless channel modelling and fading mitigation techniques, and in the third part we discuss the possible application of advanced digital signal processing, especially time-frequency representation and blind source separation, to wireless communication systems. The first part considers general digital communication systems which will be incorporated in later parts. Today's wireless communication system is a subbranch of a general digital communication system that employs various techniques of A/D (Analog to Digital) conversion, source coding, error correction, coding, modulation, and synchronization, signal detection in noise, channel estimation, and equalization. We study and develop the digital communication algorithms to enhance the performance of wireless communication systems. In the Second Part we focus on wireless channel modelling and fading mitigation techniques. A modified Jakes' method is developed for Rayleigh fading channels. We investigate the level-crossing rate (LCR), the average duration of fades (ADF), the probability density function (PDF), the cumulative distribution function (CDF) and the autocorrelation functions (ACF) of this model. The simulated results are verified against the analytical Clarke's channel model. We also construct frequency-selective geometrical-based hyperbolically distributed scatterers (GBHDS) for a macro-cell mobile environment with the proper statistical characteristics. The modified Clarke's model and the GBHDS model may be readily expanded to a MIMO channel model thus we study the MIMO fading channel, specifically we model the MIMO channel in the angular domain. A detailed analysis of Gauss-Markov approximation of the fading channel is also given. Two fading mitigation techniques are investigated: Orthogonal Frequency Division Multiplexing (OFDM) and spatial diversity. In the Third Part, we devote ourselves to the exciting fields of Time-Frequency Analysis and Blind Source Separation and investigate the application of these powerful Digital Signal Processing (DSP) tools to improve the performance of wireless communication systems.

The quadratic class of time-frequency distributions (TFDs) forms a set of tools which allow to effectively extract important information from a nonstationary signal. To determine which TFD best represents the given signal, it is a common practice to visually compare different TFDs' time-frequency plots, and select as best the TFD with the most appealing plot. This visual comparison is not only subjective, but also difficult and unreliable especially when signal components are closely-spaced in the time-frequency plane. To objectively compare TFDs, a quantitative performance measure should be used. Several measures of concentration/complexity have been proposed in the literature. However, those measures by being derived with certain theoretical assumptions about TFDs are generally not suitable for the TFD selection problem encountered in practical applications. The non-existence of practically-valuable measures for TFDs' resolution comparison, and hence the non-existence of methodologies for the signal optimal TFD selection, has significantly limited the use of time-frequency tools in practice. In this thesis, by extending and complementing the concept of spectral resolution to the case of nonstationary signals, and by redefining the set of TFDs' properties desirable for practical applications, we define an objective measure to quantify the quality of TFDs. This local measure of TFDs' resolution performance combines all important signal time-varying parameters, along with TFDs' characteristics that influence their resolution. Methodologies for automatically selecting a TFD which best suits a given signal, including real-life signals, are also developed. The optimisation of the resolution performances of TFDs, by modifying their kernel filter parameters to enhance the TFDs' resolution capabilities, is an important prerequisite in satisfying any additional application-specific requirements by the TFDs. The resolution performance measure and the accompanying TFDs' comparison criteria allow to improve procedures for designing high-resolution quadratic TFDs for practical time-frequency analysis. The separable kernel TFDs, designed in this way, are shown to best resolve closely-spaced components for various classes of synthetic and real-life signals that we have analysed.

La nécessaire représentation en échelle du monde nous amène à expliquer pourquoi la théorie des ondelettes en constitue le formalisme le mieux adapté. Ses performances sont comparées à d'autres outils : la méthode des étendues normalisées (R/S) et la méthode par décomposition empirique modale (EMD).La grande diversité des bases analysantes de la théorie des ondelettes nous conduit à proposer une approche à caractère morphologique de l'analyse. L'exposé est organisé en trois parties.Le premier chapitre est dédié aux éléments constitutifs de la théorie des ondelettes. Un lien surprenant est établi entre la notion de récurrence et l'analyse en échelle (polynômes de Daubechies) via le triangle de Pascal. Une expression analytique générale des coefficients des filtres de Daubechies à partir des racines des polynômes est ensuite proposée.Le deuxième chapitre constitue le premier domaine d'application. Il concerne les plasmas de bord des réacteurs de fusion de type tokamak. Nous exposons comment, pour la première fois sur des signaux expérimentaux, le coefficient de Hurst a pu être mesuré à partir d'un estimateur des moindres carrés à ondelettes. Nous détaillons ensuite, à partir de processus de type mouvement brownien fractionnaire (fBm), la manière dont nous avons établi un modèle (de synthèse) original reproduisant parfaitement la statistique mixte fBm et fGn qui caractérise un plasma de bord. Enfin, nous explicitons les raisons nous ayant amené à constater l'absence de lien existant entre des valeurs élevées du coefficient d'Hurst et de supposées longues corrélations.Le troisième chapitre est relatif au second domaine d'application. Il a été l'occasion de mettre en évidence comment le bien-fondé d'une approche morphologique couplée à une analyse en échelle nous ont permis d'extraire l'information relative à la taille, dans un écho rétrodiffusé d'une cible immergée et insonifiée par une onde ultrasonore / The necessary scale-based representation of the world leads us to explain why the wavelet theory is the best suited formalism. Its performances are compared to other tools: R/S analysis and empirical modal decomposition method (EMD). The great diversity of analyzing bases of wavelet theory leads us to propose a morphological approach of the analysis. The study is organized into three parts. The first chapter is dedicated to the constituent elements of wavelet theory. Then we will show the surprising link existing between recurrence concept and scale analysis (Daubechies polynomials) by using Pascal's triangle. A general analytical expression of Daubechies' filter coefficients is then proposed from the polynomial roots. The second chapter is the first application domain. It involves edge plasmas of tokamak fusion reactors. We will describe how, for the first time on experimental signals, the Hurst coefficient has been measured by a wavelet-based estimator. We will detail from fbm-like processes (fractional Brownian motion), how we have established an original model perfectly reproducing fBm and fGn joint statistics that characterizes magnetized plasmas. Finally, we will point out the reasons that show the lack of link between high values of the Hurst coefficient and possible long correlations. The third chapter is dedicated to the second application domain which is relative to the backscattered echo analysis of an immersed target insonified by an ultrasonic plane wave. We will explain how a morphological approach associated to a scale analysis can extract the diameter information