Arturas Serackis

Dynamic adaptation of the jitter buffer for video streaming applications

Paper presents a new algorithm proposed to dynamically adapt jitter buffer, used in video transcoding applications. When the video content provider is moving continuously, the wireless network connectivity conditions changes in time. The resulting jitter should be compensated in video transcoder by selecting jitter buffer high enough to compensate all possible jitter variations. The algorithm, proposed in this paper updates the jitter buffer according to current jitter measurements and jitter prediction results in order to reduce the total latency of the video transcoding in live transmissions.

A new approach for spectrum sensing in wideband

A new algorithm of the spectrum sensing is proposed in this paper. The spectrum sensing methods presented in this paper are optimised to implement these in FPGA based embedded systems. The low power and highly parallelised architecture of FPGA requires low complexity in implementation of separate processing units - spectrum sensors (SpS). The widely used detection methods, based on analysis of signal spectrum energy and standard deviation are integrated in proposed spectrum analysis algorithm for cognitive radio spectrum sensing applications. The experimental investigation of proposed algorithm is performed in simulated and real RF environments. It is showed that proposed algorithm increase the spectrum sensing efficiency and minimise the miss-detection of licensed users to 12 % using only the energy detector in proposed algorithm and 5% additionally adding the standard deviation based spectrum sensing. Three channel verification techniques with proposed spectrum sensors were tested during our investigation. The Forward Consecutive Mean Excision (FCME) based technique showed the highest accuracy in channel verification task.

Word Recognition Acceleration by Double Random Seed Matching in Perceptual Cepstrum Error Space

Paper presents an algorithm for acceleration of the dynamic time warping (DTW) based isolated word recognition algorithm. The number of matching operations directly depends on the size of vocabulary. A set of perceptual cepstrum features is calculated for each word and stored in the vocabulary as a reference. Additionally all words (references) are compared between each other using DTW in order to get the reference-to-reference matches. The acceleration of pattern matching is acquired by adaptive search of the pattern reference according to the previous matching results ant reference-to-reference matches. A modified word selection scenario applied for the vocabulary reduces the number of matching operations by 62-70 % in average. The reduction of matching operations allows to use DTW based speech recognition methods in real-time control applications and only need additional 13 % of vocabulary storage space.

Prediction of the real-time video streaming performance based on the peer connection statistics

The aim of the investigation presented in this paper was to design a video stream performance predictor, which could be used for adaptive video streaming applications. The neural networks based predictors were analyzed in this paper. An experimental investigation was performed in order to test the once trained predictors on a real WebRTC statistical data, recorded in dynamically changing mobile data throughput conditions.

Application of k-means and MLP in the automation of matching of 2DE gel images

Critical information that is related to vital processes of the cell can be revealed comparing several two-dimensional electrophoresis (2DE) gel images. Through up to 10 000 protein spots may appear in inevitably noisy gel thus 2DE gel image comparison and analysis protocols usually involve the work of experts. In this paper we demonstrate how the problem of automation of 2DE gel image matching can be gradually solved by the use of artificial neural networks. We report on the development of feature set, built from various distance measures, selected and grounded by the application of self-organizing feature map and confirmed by expert decisions. We suggest and experimentally confirm the use of k-means clustering for the pre-classification of 2DE gel image into segments of interest that about twice speed-up the comparison procedure. We develop original Multilayer Perceptron based classifier for 2DE gel image matching that employs the selected feature set. By experimentation with the synthetic, semi-synthetic and natural 2DE images we show its superiority against the single distance metric based classifiers.

A Robust Identification of the Protein Standard’s Bands in Two-Dimensional Electrophoresis Gel Images

Abstract The aim of the investigation presented in this paper was to develop a software-based assistant for the protein analysis workflow. The prior characterization of the unknown protein in two-dimensional electrophoresis gel images is performed according to the molecular weight and isoelectric point of each protein spot estimated from the gel image before further sequence analysis by mass spectrometry. The paper presents a method for automatic and robust identification of the protein standard band in a two-dimensional gel image. In addition, the method introduces the identification of the positions of the markers, prepared by using pre-selected proteins with known molecular mass. The robustness of the method was achieved by using special validation rules in the proposed original algorithms. In addition, a self-organizing map-based decision support algorithm is proposed, which takes Gabor coefficients as image features and searches for the differences in preselected vertical image bars. The experimental investigation proved the good performance of the new algorithms included into the proposed method. The detection of the protein standard markers works without modification of algorithm parameters on two-dimensional gel images obtained by using different staining and destaining procedures, which results in different average levels of intensity in the images.

The prediction of cut-off frequencies of models of gyroelectric waveguides using artificial neural networks

Gyroelectric n and p types waveguides were usually investigated using differential Maxwells equations, coupled mode and partial area methods, coherent approaching, least square methods. Computation time of one particular model of this type of waveguide might take quite a long time and even to a couple of days using one of these analytical methods. The whole investigation may require a lot of time in the first stage of research until the right model of waveguide will be found. The artificial neural networks were adjusted for the investigation of gyroelectric n GaAs waveguides. Multilayer perceptron network was selected during investigation. Advantages of artificial neural networks comparing with analytical methods are presented in this paper. The investigation showed that difference between results, obtained using analytical methods, and results, obtained by using artificial neural networks, do not differ by more than 12%. On the other hand prediction using artificial neural networks is performed about 2000 times faster than using traditional methods.

A new self-organizing map topology for real-time spectrum sensing and fast convergence

The paper presents a new modification of a self-organizing map topology to increase the speed of feature map convergence. The modification proposed in this paper is useful for real time spectrum sensing applications in which the SOM is used for classification tasks. Three alternative SOM topologies were tested and compared with the proposed topology modification. The number of epochs needed to ensure the convergence of feature maps was reduced by about 20 times using the proposed topology modification. An experimental investigation was performed using an SOM-based spectrum sensor on two radio bands with different spectral characteristics. The modified SOM topology showed not only a faster convergence, but also a better emission detection performance comparing to hexagonal, grid and rhombus topologies.

Identification of ECG signal pattern changes to reduce the incidence of Ventricular Tachycardia false alarms

The paper focuses on the reduction of the false alarms in the Intensive Care Units (ICU). Five alarm types were analyzed in this study: Asystole, Extreme Bradycardia, Extreme Tachycardia, Ventricular Tachycardia and Ventricular Flutter/Fibrillation. Most of the analyzed alarm types rely on the quality of the heart rate estimation. The false alarm reduction algorithms analyzed in this paper use the quality estimate of the arterial blood pressure signal from which the heart rate is estimated and additionally the results of heart beat detection in two ECG signals are analyzed before making the final decision about the true or false alarm type. The most attention in this paper is focused on the correct detection of Ventricular Tachycardia alarms. The decision about the true or false alarm is made according to RR interval variation and changes of QRS complex shape features. A subset of sample entries data of the Physionet/CinC Challenge 2015 is used to test the proposed algorithm modifications. The false alarm detection according to the RR interval variation gave 49% TPR, 49% TNR (score 34.82) for the Phase I Entries data set and 46% TPR, 51% TNR (score 34.59) for the Phase II Entries data set. The VT alarm detection algorithm based on the features related to the the ECG waveform shape has increased the VT score for Phase I Entries data set to 41.98.