The aim of this work is to create a web application for automatic topic detection and segmentation of on-line courses. During playback of processed records, the application should be able to offer records from thematically consistent on-line courses. This document contains problem description, list of used instruments, description of implementation, the principle of operation and description of final user interface.

The Convolutional Neural Network (CNN) is a mathematical model designed to distill input information into a more useful representation. This distillation process removes information over time through a series of dimensionality reductions, which ultimately, grant the model the ability to resist noise, and generalize effectively. However, CNNs often contain elements that are ineffective at contributing towards useful representations. This Thesis aims at providing a remedy for this problem by introducing Absolute Cosine Value Regularization (ACVR). This is a regularization technique hypothesized to increase the representational power of CNNs by using a Gradient Descent Orthogonalization algorithm to force the vectors that constitute their filters at any given convolutional layer to occupy unique positions in Rⁿ. This method should in theory, lead to a more effective balance between information loss and representational power, ultimately, increasing network performance. The following Thesis proposes and examines the mathematics and intuition behind ACVR, and goes on to propose Dynamic-ACVR (D-ACVR). This Thesis also proposes and examines the effects of ACVR on the filters of a low-dimensional CNN, as well as the effects of ACVR and D-ACVR on traditional Convolutional filters in VGG-19. Finally, this Thesis proposes and examines regularization of the Pointwise filters in MobileNetv1.

The recent uprise of end-to-end speech translation models requires a new generation of parallel corpora, composed of a large amount of source language speech utterances aligned with their target language textual translations. We hereby show a pipeline and a set of methods to collect hundreds of hours of English audio-book recordings and align them with their Italian textual translations, using exclusively public domain resources gathered semi-automatically from the web. The pipeline consists in three main areas: text collection, bilingual text alignment, and forced alignment. For the text collection task, we show how to automatically find e-book titles in a target language by using machine translation, web information retrieval, and named entity recognition and translation techniques. For the bilingual text alignment task, we investigated three methods: the Gale–Church algorithm in conjunction with a small-size hand-crafted bilingual dictionary, the Gale–Church algorithm in conjunction with a bigger bilingual dictionary automatically inferred through statistical machine translation, and bilingual text alignment by computing the vector similarity of multilingual embeddings of concatenation of consecutive sentences. Our findings seem to indicate that the consecutive-sentence-embeddings similarity computation approach manages to improve the alignment of difficult sentences by indirectly performing sentence re-segmentation. For the forced alignment task, we give a theoretical overview of the preferred method depending on the properties of the text to be aligned with the audio, suggesting and using a TTS-DTW (text-to-speech and dynamic time warping) based approach in our pipeline. The result of our experiments is a publicly available multi-modal corpus composed of about 130 hours of English speech aligned with its Italian textual translation and split in 60561 triplets of English audio, English transcript, and Italian textual translation. We also post-processed the corpus so as to extract 40-MFCCs features from the audio segments and released them as a data-set.

Coherent plane-wave compounding (CPWC) ultrasound is an important modality enabling ultrafast biomedical imaging. To perform CWPC image reconstruction for a stratified (horizontally layered) medium, one needs to know how the speed of sound (SOS) varies with the propagation depth. Incorrect sound speed and layer thickness assumptions can cause focusing errors, degraded spatial resolution and significant geometrical distortions resulting in poor image reconstruction. We aim to determine the speed of sound and thickness values for each horizontal layer to accurately locate the recorded reflection events to their true locations within the medium. Our CPWC image reconstruction process is based on phase-shift migration (PSM) that requires the user to specify the speed of sound and thickness of each layer in advance. Prior to performing phase-shift migration (one layer at a time, starting from the surface), we first estimate the speed of sound values of a given layer using a cosine similarity metric, based on the data obtained by a multi-element transducer array for two different plane-wave emission angles. Then, we use our speed estimate to identify the layer thickness via end-of-layer boundary detection. A low-cost alternative that obtains reconstructed images with fewer phase shifts (i.e., fewer complex multiplications) using a spectral energy threshold is also proposed in this thesis. Our evaluation results, based on the CPWC imaging simulation of a three-layer medium, show that our sound speed and layer thickness estimates are within 4% of their true values (i.e., those used to generate simulated data). We have also confirmed the accuracy of our speed and layer thickness estimation separately, using two experimental datasets representing two special cases. For speed estimation, we used a CPWC imaging dataset for a constant-speed (i.e., single-layer) medium, yielding estimates within 1% of their true values. For layer thickness estimation, we used a monostatic (i.e., single-element) synthetic-aperture (SA) imaging dataset of the three-layer medium, also yielding estimates within 1% of their true values. Our evaluation results for the low-cost alternative showed a 93% reduction in complex multiplications for the three-layer CPWC imaging dataset and 76% for the three-layer monostatic SA imaging dataset, producing images nearly similar to those obtained using the original PSM methods. / Graduate

We describe a fast method for the evaluation of an arbitrary high-dimensional multivariate algebraic polynomial in Chebyshev form at the nodes of an arbitrary rank-1 Chebyshev lattice. Our main focus is on conditions on rank-1 Chebyshev lattices allowing for the exact reconstruction of such polynomials from samples along such lattices and we present an algorithm for constructing suitable rank-1 Chebyshev lattices based on a component-by-component approach. Moreover, we give a method for the fast, exact and stable reconstruction.

No description available.

No description available.

The task of detecting human faces within either a still image or a video frame is one of the most popular object detection problems. For the last twenty years researchers have shown great interest in this problem because it is an essential pre-processing stage for computing systems that process human faces as input data. Example applications include face recognition systems, vision systems for autonomous robots, human computer interaction systems (HCI), surveillance systems, biometric based authentication systems, video transmission and video compression systems, and content based image retrieval systems. In this thesis, non-traditional methods are investigated for detecting human faces within color images or video frames. The attempted methods are chosen such that the required computing power and memory consumption are adequate for real-time hardware implementation. First, a standard color image database is introduced in order to accomplish fair evaluation and benchmarking of face detection and skin segmentation approaches. Next, a new pre-processing scheme based on skin segmentation is presented to prepare the input image for feature extraction. The presented pre-processing scheme requires relatively low computing power and memory needs. Then, several feature extraction techniques are evaluated. This thesis introduces feature extraction based on Two Dimensional Discrete Cosine Transform (2D-DCT), Two Dimensional Discrete Wavelet Transform (2D-DWT), geometrical moment invariants, and edge detection. It also attempts to construct a hybrid feature vector by the fusion between 2D-DCT coefficients and edge information, as well as the fusion between 2D-DWT coefficients and geometrical moments. A self organizing map (SOM) based classifier is used within all the experiments to distinguish between facial and non-facial samples. Two strategies are tried to make the final decision from the output of a single SOM or multiple SOM. Finally, an FPGA based framework that implements the presented techniques, is presented as well as a partial implementation. Every presented technique has been evaluated consistently using the same dataset. The experiments show very promising results. The highest detection rate of 89.2% was obtained when using a fusion between DCT coefficients and edge information to construct the feature vector. A second highest rate of 88.7% was achieved by using a fusion between DWT coefficients and geometrical moments. Finally, a third highest rate of 85.2% was obtained by calculating the moments of edges. / Master of Science

Events profoundly influence human-environment interactions. Through repetition, some events manifest and amplify collective behavioral traits, which significantly affects landscapes and their use, meaning, and value. However, the majority of research on reaction to events focuses on case studies, based on spatial subsets of data. This makes it difficult to put observations into context and to isolate sources of noise or bias found in data. As a result, inclusion of perceived aesthetic values, for example, in cultural ecosystem services, as a means to protect and develop landscapes, remains problematic. In this work, we focus on human behavior worldwide by exploring global reactions to sunset and sunrise using two datasets collected from Instagram and Flickr. By focusing on the consistency and reproducibility of results across these datasets, our goal is to contribute to the development of more robust methods for identifying landscape preference using geo-social media data, while also exploring motivations for photographing these particular events. Based on a four facet context model, reactions to sunset and sunrise are explored for Where, Who, What, and When. We further compare reactions across different groups, with the aim of quantifying differences in behavior and information spread. Our results suggest that a balanced assessment of landscape preference across different regions and datasets is possible, which strengthens representativity and exploring the How and Why in particular event contexts. The process of analysis is fully documented, allowing transparent replication and adoption to other events or datasets.

This doctoral thesis is focused on problems of modulators based on six-ports. It begins with description of current state of the art of six-ports used like modulators, their transfer functions and SIW technology. A design part of this thesis consists from experimental six-port based on substrate integrated waveguide (SIW) technology. There is presented step-by-step development of this six-port using this technology and also there is introduced micro-strip technology based six-port. Final design of six-ports and variable impedances were measured, the results are discussed and compared with expected ones in next chapters. Second part of this thesis deals with influences of internal parameters of six-ports to final signal transmission and derives theirs transfer functions for more than one reflection in structure. The computation results are compared with experimental measurements for fixed loads. With using of ideal loads sweeps, modulations with shaped input signals were calculated. For designed variables impedances, there was founded the optimal biasing points for demanded IQ diagram and discussed which from tested active circuit is suitable. In the last part there are shown results of experiment with these variable loads connected to both types of designed six-ports.