CC BY
7Towards real time fluid motion estimation by particle images based on convolutional neural networks
W. Koliai1, M. Tokarev1,2
1Novosibirsk State University
2Institute of Thermophysics SB RAS
Email: mtokarev@itp.nsc.ru
DOI 10.24412/cl-35065-2021-1-02-64
A task of motion estimation can be found in various fields from surveillance security systems, industrial in-
spection and non-destructive testing to computer graphics increasing rendering speed and visual motion cap-
ture. Additionally whole-field quantitative assessment of mechanical stresses and velocity in the strain gauges
and fluid velocity sensors can be mentioned. Optical flow processing algorithms are frequently applied when
using dense motion estimation [1]. This technique is based on conservation of local intensity to track dis-
placements of moving objects within the image [1]. Two-dimensional motion estimation for fluids is usually
performed with high power pulsed laser illumination, and it is difficult to guarantee similar intensity profiles
between pulses. Developing alternative fast motion estimation algorithms that can be adjusted for specific
intensity variation conditions is relevant. This work presents test results of application of a convolutional neu-
ral network model PWC-Net [2] for on-line estimation of dense velocity field using pairs of particle images of
VGA size. The original model was fine-tuned by 15 k synthetic moving particle image pairs [3] modeling differ-
ent turbulent flow configurations as well as additional 1 k real experimental images. Currently achieved infer-
ence time as low as 70 ms (14FPS) can be improved further modifying the model configuration sacrificing spa-
tial resolution and accuracy of the results.
This work was supported by the state contract with IT SB RAS.
References
1. Horn B. K. P., Schunck B. G. Determining optical flow // Artif. Intell. 1981. V. 17, P. 185�203.
2. Sun D., Yang X., Liu M. Y., Kautz J. PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume //
Proceedings of the IEEE conference on computer vision and pattern recognition. 2018. P. 8934-8943.
3. Cai S., Zhou S., Xu C., Gao Q. Dense motion estimation of particle images via a convolutional neural network //
Experiments in Fluids. 2019. V. 60(4), P. 1-16.
The use of neural networks for predicting concentration limits of flammability
O. V. Krivetchenko
Novosibirsk State University of Economics and Management
Email: kriv_ok@ngs.ru
DOI 10.24412/cl-35065-2021-1-02-65
Neural networks were used to predict the flammability�s concentration limits of chemicals of various clas-
ses, based on a combination of descriptor and neural network approaches. Fragmentary descriptors are used
in forecasting. They are calculated automatically using the structural formula of substance. It was detected
that the results obtained in the article for predicting the concentration limits of flammability make an insignifi-
cant error in comparison to the experiment.
For a comparative analysis of the developed methods capabilities, the equations of domestic and foreign
researchers are taken.
The research suggests and studies the methods combining the neural network and descriptor approaches
for predicting the concentration limits of chemicals flammability. The best data are found using the �atom-
bond-atom� descriptors.
References
1. Osipov L.A. Prediction of flammability�s concentration limits on the basis of information and charge descriptors /
A.L. Osipov, O.V. Krivetchenko, V.P. Trushina // International J. of advanced studies. 2018. Vol. 8. No. 4-2. P. 83-91.
Selective regularization of linear regression model
V. N. Lutay, N. S. Khusainov
Southern Federal University, Rostov-on-Don
Email: vnlutay@sfedu.ru
DOI 10.24412/cl-35065-2021-1-02-66
The paper considers the construction of a linear regression model with regularization of the matrix of the
system of normal equations. Unlike traditional ridge regression, which consists in adding some positive pa-
rameter to all diagonal elements of the matrix [1], the proposed method increases only those diagonal ele-
ments that help to reduce the condition number of the matrix and, consequently, reduce the variance of the
coefficients of the regression equation. These are the smallest diagonal elements of the triangular matrix ob-
tained by the Cholesky decomposition of the correlation matrix of the original data set [2]. The effectiveness of
the method is tested on a known dataset, and the comparison is made not only with the ridge regression, but
also with the results of the widely used algorithms Lasso and LARS [3].
References
1. Draper N., Smith H. Applied Regression Analysis, 2 Edition. John Wiley & Sons 2007.
2. Lutay V.N. Improving the Stability of Triangular Decomposition of Ill-Conditioned Matrices // Numerical Analysis
and Applications. 2019. V.12, N.4. P. 388�94.
3. Efron B, Hastie T, Johnstone J and Tibshirani R. Least Angle Regression// The Annals of Statistics. 2004. V. 32.
P. 407-499.
Researching zero-shot learning methods for automatic speech recognition
A. Yu. Mikhaylenko, I. Yu. Bondarenko
Novosibirsk State University
Email: mikkhailenko@gmail.com, i.bondarenko@g.nsu.ru
DOI 10.24412/cl-35065-2021-1-02-67
Speech recognition systems require datasets that will contribute to high quality recognition. However,
such systems often do not meet this characteristic due to the lack of labeled data. The lack of labeled data, in
turn, entails an imbalance: the high accuracy of the model will turn out to be "false" � the input data will be
biased towards a particular class, which will affect the trained model. We propose a new in speech recognition
termins method of dealing with this problem: training generative models [1, 2] on objects of one set of classes
(seen) and classifying objects of another set of classes (unseen) using additional information. Using generative
models helps us generate signal features for invisible classes using semantic information. A generative net-
works generates signal characteristics based on the semantic attributes of a particular class.
References
1. Schonfeld E., Ebrahimi S., Sinha S., Darrell T., Akata Z. �Generalized Zero- and Few-Shot Learning via Aligned
Variational Autoencoders� // IEEE Conference on Computer Vision and Pattern Recognition. 2019. P. 1-9.
2. Xian Y., Lorenz T., Schiele B., Akata Z. Feature Generating Networks for Zero-Shot Learning // ArXiv. 2018. P. 1-10.
