Section 4 - Track 4

On this page, talks in the format of posters are posted. They can be viewed during the whole time of the Conference. To send your question to the authors, please follow to Miro and write your question in the comments. You can also use the feedback form on our website. Authors will answer questions using Miro comments or via email. To get manual go to Instructions (eng / rus).

Vladimir Chernov
A new approach to the synthesis of parallel error-free computing systems
A key role in ensuring the reliability of hydraulic systems is operational reliability, which is achieved by using modern diagnostic methods. Existing methods for monitoring the technical condition of the hydraulic system during their operation in most cases are based on achieving the required indicators for the output parameters, such as pressure, flow rate of the working fluid, the speed of shifting of the actuators, leakage and other parameters calculated by the duration of the transient processes. This approach allows you to diagnose the overall hydraulic systems and its components in detail. This paper describes an array of data generation and preparation and real-time diagnosis on the operation for a working and faulty hydraulic system in order to develop a tool for operational and proactive maintenance of industrial plants and technological complexes in accordance with their technical condition. Experimental test bench is used for data preparation and online tests of machine learning, neural network fault detection and prediction methods.

Igor Rytsarev
Text data analysis using conversion analysis
In this paper, we propose an algorithm for analyzing text data based on conversion analysis. Currently, natural languages are developing dynamically. New semantic units constantly come into spoken language. Under these conditions, chains of dependency graphs of semantic units are constantly being rebuilt. In this paper, we propose a method for determining synonyms based on conversion analysis. The proposed method was tested on data that was collected from social networks

Dmitry Ulyanov and Dmitry Savelyev
The investigation of the using the cyclic generative-competitive neural networks for image stylization
The paper provides examples of convolutional neural network architectures, the corresponding activation functions, and the organization of their interaction in the learning process. Networks interact with each other according to the architecture of generative-adversarial networks. For the task, the NEXET 2017 data set was filtered and formatted. Studies of the architecture of neural networks and varying the volume of the training sample to solve the problem of image styling were carried out.

Daria Arkhipova and Egor Goshin
Influence of image set formation parameters on the result of super-resolution reconstruction
This article is dedicated to the study of influence of image sequence formation parameters and algorithm parameters on the result of reconstruction of high-resolution images using the projection on convex sets method. For the parameters of the sequence formation the image reduction scale along both axes and the number of images in the sequence were selected. For the algorithm parameters we used the admissible value of the residual during the formation of convex sets and the window size of the point spread function. To set up an experiment, the projection method on convex sets was implemented in the Python programming language. The result of experiments on a test set of images is presented.

Sergey Parkhomenko
The creation of SDN testbed for network security algorithms development
This paper describes a testbed for network security algorithms development using the capabilities of software-defined networks (SDN). Structural chart of testbed and description of its modules for processing NetFlow data and restricting malicious traffic from attacking IP address are given. The approach for detecting network anomalies based on the determination of threshold values for network variables is described.

Marina Golovastikova and Andrey Gaidel
Texture images сlassification using deep learning techniques
The aim of the work is to study the reliability of the texture images classification depending on the architecture of the neural network and input data. In the course of the work, a network architecture was proposed that receives texture images as an input and allows achieving a classification reliability of 92%. As an alternative, a fully-connected neural network was proposed, which receives an input vector of image features, which, while saving time by 48 times, shows a result that is 8% less.

Yegor Goshin, Maksim Marchukov and Anton Kotov
Information technology for image stitching
In this article we study two ways to create mosaic from images. First way is based on finding feature points on the images and aligning of them, second is based on Fourier-Mellin transform. This paper shows application areas for both ways and propose adaptive method for choosing better way for image stitching based on image’s characteristics.

Nickolay Shlyankin and Andrey Gaidel
Application of the Hidden Markov Model for determining PQRST complexes in electrocardiograms
The paper considers the main aspects of ECG signal segmentation, estimates of the methods used are obtained, and their advantages and disadvantages are examined. The application of hidden Markov models with various parameters for segmenting QRS, ST, T, P, PQ, ISO complexes of electrocardiograms is investigated. QT database. For comparison, the Pan-Tompkins algorithm for searching for the duration of QRS complexes was modified.

Kirill Musin and Andrey Gaidel
Machine learning algorithms in the prediction of conflicts in clinical classification of genetic variants
The clinical classification of a person’s genetic variant can lead to conflicting classifications. The presence of conflicts is determined manually by laboratory methods. If there is a conflict, then there is a difficulty in interpreting the result. In this work, with the help of machine learning algorithms, it was possible to train the neural network to predict conflicts with an accuracy of 78%, and also to determine which parameters are most important in the classification.

Alexandr Rud, Sergey Rud, Michael Isaev and Dmitry Savelyev
The using convolutional neural networks for determine the age of a person from an image
The paper presents the results of research to determine the biological age of a person using the image of the face. To solve this problem, we used a random forest algorithm using a hybrid Hesse filter and a local binary template operator, as well as a convolutional neural network ResNet50. The data sets used, the problems associated with their application, as well as the accuracy of classification in the selected division of the age line are presented.

Igor Kilbas and Rustam Paringer
Gradient as a foundation for building a loss function
Deep neural networks have achieved a tremendous success in a variety of applications across many disciplines. Yet, training a neural network comes with several challenges that have to be solved. The performance of a deep learning models rely not only on the network architecture but also on the choice of a loss function. Cross-entropy loss found to be the most common choice for the classification problem. But it's main downside is that it can't handle data with huge class imbalance. In order to tackle this problem the Focal loss has been proposed. In this paper we investigate the reasons behind its good performance. We find several properties of the Focal loss' gradient that can be applied for building new loss functions and propose a few of them. We also show an experimental evidence of the validity of the proposed functions.

Stanislav Abulkhanov
Structural changes in microroughness preceding surface fatigue failure
The approach of controlling small dynamic changes in the image of the process under study using cyclic codes is considered. We performed a comparative analysis of the same pixel sequences in the original image and the corresponding pixel sequences in the monitored image. Sequences of identical pixels we considered as bits of the original message. The generating polynomials of cyclic codes corresponding to the initial and controlled image, in our opinion, characterize the state of the process under study at different points in time. We interpret the difference between the generating polynomials as an external influence (interference), which leads to changes on the controlled surface. To visualize the interference, we used the packaging method.

Elizaveta Rudinskaya and Rustam Paringer
Study of a face detection accuracy based on race and gender using Haar cascades
In this work, we examine the problem of face detection in images based on the gender and the race of a person. The paper focuses on the search of the detection parameters, creating a training dataset and obtaining results with greater face detection accuracy.

Stanislav Abulkhanov, Ivan Bayrikov, Dmitriy Goryainov and Oleg Slesarev
Titanium cellular implant to replace bone defects in the jaw
The article suggests the technique of designing a titanium graft with cellular structure. The size of the cells provides weight and rigidity of the construction, which corresponds to the weight and rigidity of the bone tissue.

Yegor Goshin, Pavel Volkhin and Anton Kotov
Research of formats of test data sets for solving the odometry problem
This article provides an overview of various sets of test data for conducting researches to assess the quality of algorithms for determining the movement and position of a camera from a sequence of optical images. Based on the review, the task of presenting data in a given format is formulated. To solve this problem, a software tool has been developed for converting the formats for representing the parameters of the initial test suites in the Python.

Ilya Smirnov, Igor Rytsarev, Alexander Kupriyanov and Dmitriy Kirsh
Development of algorithms for annotating information in social networks
The work is devoted to the research of algorithms for annotating information in social networks. As the object of research, the social network Instagram was selected. To solve the problem of obtaining the necessary information, studies in the field the data collection of the social network Instagram were carried out. A software tools that provides the collection of necessary data and the annotates information have been developed. The existing algorithms for annotating information have been investigated and mined on images as basic data from an Instagram social network.

Evgeniy Minaev
High performance implementation of machine learning method based on fractal compression
This article investigated the machine learning method based on fractal compression. The main idea is to apply the fractal compression method based on iterated function systems to reduce the dimension of the original images. High performance implementation of machine learning method based on fractal compression is proposed in this article.

Rail Gabbasov and Rustam Paringer
Influence of the receptive field size on accuracy and performance of a convolutional neural network
Convolutional neural networks (CNNs) have been successfully applied to many tasks such as digit and object recognition. In this paper we study the size of receptive field of deep convolutional neural networks, in particular, we check the idea of a “redundant” receptive field. We run a set of experiments on two common CNN models — VGG16 and ResNet18 — in order to explore the influence of receptive field size on CNN's training time, accuracy and performance. We run experiments using MakiFlow framework on CALTECH256 dataset. The experiments' results show that optimization of neural networks (NNs) by reducing the size of receptive field allows to reduce the NN’s training time by 5-7% while maintaining the accuracy of the network.

Sergey Kostin and Andrey Gaidel
Predicting exchange rate dynamics in the forex market using machine learning
The paper investigates the effectiveness of two machine learning algorithms for predicting the exchange rate. The first algorithm is based on linear regression, and the second is based on a recurrent neural network with long short-term memory. The main advantage of these algorithms is that they are well suited for predicting time series. To train and build a regression model, the sliding window method is used, which allows you to use the previous time steps to predict the next step. As an assessment of the obtained models, the coefficient of determination, the standard error of the regression losses, and the average absolute error of the regression losses are used. Both algorithms show comparable results.

Sergei Stepanenko and Pavel Yakimov
Development of a cloud platform for gathering, storing and analysis of video data
Today video analysis is very relevant. Amount of video is being increased and there is a need to perform it. Very often there are a number of video data sources on which video is being continuosly recorded. This article presents an implementation of a platform which aggregates video from various sources and processes it with use of different services which are based on artificial neural networks and machine learning methods.

Pavel Katkov and Alexander Khramov
Detection of blackout in the lungs by X-ray DICOM image using neural networks
According to statistics, in Russia in 2018, the existing number of people died due to respiratory and respiratory diseases: there were 59,803 of them. Almost half of these deaths (24954) came from pneumonia. One of the signs of pneumonia is blackout in the lungs. This study shows that this can be detected in an x-ray image using neural networks. During the research, a convolutional neural network was implemented, which takes place if it exists. The accuracy of the neural network is estimated using the average values of the Jacquard coefficients (0.71).

Artem Gaidar, Pavel Yakimov and Andrey Viktorenkov
Identification of defects in the inside of a metal pipe
This paper aims to provide an overview of the experimental and simulation works focused on the detection, localisation and assessment of various defects in pipes by applying computer vision techniques that have been used in the oil and gas industries over the past 20 years. Designed by software package for the detection and classification a metal surface defects using a artificial intelligence.

Ekaterina Avdonina and Pavel Yakimov
Research of algorithm of detection of a pose of the person on the image and in a video stream
This article is devoted to the study of various algorithms for detecting a person in an image and in a video stream. Comparison of algorithms based on calculation accuracy using datasets. The choice of the algorithm with the most accurate results and the consumption of the least amount of resources on devices with different technical characteristics.

Dmitriy Kirsh
Parametric Identification of Crystal Lattices Based on Isosurface Configuration Analysis
The paper deals with an approach to crystal lattice identification based on the isosurface configuration analysis. The approach allows us to evaluate the relative position of lattice nodes inside the unit cell. Theoretically, this should provide favourable conditions for creating identification methods that are resistant to structural distortions. The crucial problem with the high computational complexity of the algorithm is solved by modification of the method using the periodicity property of crystal lattice structures. A study of the developed method on a large base of reference lattices confirmed its high resistance to structural distortions up to a maximum distortion of 10 % of unit cell’s size.

Kirill Pugachev and Vladimir Fursov
Use of conformity principle in the visual odometry problem
In this paper we present a new proximity measure of vectors based on the principle of conformity. The measure is calculated as the sum of all possible combinations of the squared differences of the compared vectors’ elements. Our proximity measure was applied in the image matching problem using SIFT descriptors. The effectiveness of the proposed proximity measure is compared with the sum of squared differences. As a comparison criterion, we use the ratio of correctly determined corresponding cross-checked points to their total number. It was found that conforming proximity measure allows to increase a probability of correct feature point matching and uniqueness of descriptors. The results of experiments on test images are presented.

Yuriy Kurbatov, Igor Rytsarev and Alexander Kupriyanov
Research of text data processing algorithms in social networks
In this paper, we investigate various clustering algorithms for a large amount of text data. An analysis of the existing implementation methods was carried out and the algorithms Word2Vec and GloVe were selected. The initial textual data for testing the algorithms were obtained by collecting records from open VKontakte communities. The results showed that the use of these algorithms allows us to assess the frequency of use and the significance of individual words relative to the context of the studied community. The results of the algorithms` applications were compared and the conclusion about their efficiency was made in the work as well.

Andrew Galochkin and Pavel Yakimov
Development of auto-review algorithm for conference management system
This work is devoted to the creation of a conference management system with the function of automated review of submitted papers. Currently, conferences with a large audience and a variety of programs need a system for the administration of such events. In this paper we propose the system architecture of the conference management to automate time-consuming administrative tasks. An algorithm for auto-review of scientific papers is also proposed to simplify the evaluation of the conference participants ' papers. The assessment of works is based on the analysis of visual and content components. It is proposed to use convolutional neural networks for image processing.

Nikolai Skladnev and Pavel Yakimov
Development of a service for tracking the trajectory of the object when moving in room using multiple cameras
The trajectory of people moving in the context of the room is very important both for ensuring a security of a closed territory, and for social or marketing research. The main goal of this scientific research is to develop an intelligent video tracking system for an object indoors, characterized by a minimum number of errors when switching between cameras. To achieve this goal, a method of inter-camera tracking was developed using an algorithm for positioning an object on the floor plan, its main steps were considered: localizing the object in the image, projecting the object on the floor plan and combining data from several cameras.

Polina Katkova and Pavel Yakimov
3D Reconstruction via single 2D Image
Computer Vision technology is rapidly developing nowadays. The need for 3D-reconstruction methods increases along with a number of Computer Vision system implementation. The highest need is for methods, which are using single image as an input data. This article provides an overview of existing methods for 3D-reconstruction and an explanation of planned implementation, which consists of a platform and a 3D-reconstruction algorithm using single image.

Back to Top