Keynote speakers
Prof. Vikram M. Gadre
Department of Electrical Engineering, IIT Bombay, Mumbai, India
Bringing Multiresolution Signal Processing Principles into Machine Learning / Deep Learning
Abstract: Multiresolution analysis (MRA) can be considered a powerful tool, for analyzing non-stationary and transient signals. MRA has wide applications in computer vision, transient and edge detection, medical imaging, image coding/compression, de-noising, sparse signal processing, biomedical signal processing, biometrics, radar and communications, pitch estimation and geophysics. Wavelets lead to a multiresolution analysis of signals. MRA has been found to perform well in pattern recognition applications. Due to evolving technology, exponential growth in computation facilities, and availability of big databases, Deep Learning (DL) algorithms have been extensively used in pattern recognition tasks. In MRA, wavelet functions are applied to 1D or 2D signals to extract features. At first glance, MRA and DL techniques appear disjoint. However, empirically it is observed in many applications, that when both MRA and DL techniques are combined, they produce better results than the methods which employ DL techniques directly without employing MRA.
This talk will focus on the recent development in the field of MRA and applications, wherein MRA and DL can be combined, to exploit the benefits of both. In particular, the first part of the talk will discuss the recent developments of MRA along with DL techniques in touch-less biometrics for fingerprint matching, iris and ear recognition. The wavelet scattering network has been employed, with encouraging results.
In the second part of the talk, we will discuss a new flexible and versatile method, based on a deep learning architecture and augmentation technique, for a rotation invariant automatic ear recognition system. A convolutional neural network is trained with a set of manually generated data points. We have used geometric transformations that include rotation, flipping and zooming method to generate artificial data points, to increase the training images. We also explore the performance of different data augmentation techniques, for limited data ear recognition. The results are very encouraging, up to 45o rotation of ear images. The accuracy of the proposed model and their comparison with other models in terms of Model accuracy, Model loss and so on will also be discussed in the talk.
Prof. Vikram M. Gadre is working as Professor in Department of Electrical Engineering, IIT Bombay, Mumbai, India. His research interest includes Communication and signal processing, with emphasis on multi-resolution and multirate signal processing, especially wavelets and filter banks: theory and applications.
He had received President’s Gold Medal from the Indian Institute of Technology Delhi in 1989 for cumulative performance during B. Tech. He received „Excellence in Teaching‟ award from IIT Bombay in 1999, 2004, 2009 and 2014, In addition to the regular academic courses taught as a part of his duties as a faculty member at IIT Bombay, Prof V M Gadre has made a number of efforts to organize, and participate in, continuing education programmes for the industry and for other research organizations. . He has guided several sponsored research projects for organizations like Tata InfoTech, Texas Instruments etc. Prof V.M. Gadre has supervised 20 Doctoral Theses. He has authors many books in the field of Muti-resolution Analysis and wavelets.
Prof. V M Gadre is currently Principal Investigator of the „Knowledge Incubation under TEQIP‟ (TEQIP-KITE) Initiative of the Ministry of Education (GoI) at IIT Bombay, operated through the Centre for Distance Engineering Education Programme (CDEEP) and the Continuing Education Programme (CEP) in TEQIP. His two online courses Digital Signal Processing and it's Applications and Adv. Signal Processing- Multirate and Wavelets in NPTEL is quite popular among the learners of this DSP
Igor Djurović
University of Montenegro
Deep neural networks in multimedia watermarking
Abstract: A digital watermark is an invisible or inaudible signal embedded into multimedia data that can in an undisputable manner identify the author, owner, or distributor of data. Over 20 years of development, we have obtained some useful tools for this purpose. From our point of view, the watermark must be robust to changes that are not deteriorating data quality in a significant manner. Such changes are commonly referred to as attacks.
Recently, deep neural networks (DNNs) development improved results in numerous fields of signal processing. Many benchmark results are surpassed by the application of the DNNs. The DNNs are used for developing forgeries of multimedia data (deepfakes) and for breaching security in various fields. However, the current improvement of the digital watermarking techniques by DNNs is moderate. So we are now in
a paradoxical situation: the DNNs are successfully used to attack the digital watermarking and related systems, while the digital watermarking techniques are not significantly improved using these means.
In this presentation, we are going to describe the most recent development in the DNN application to digital watermarking. The developed system jointly trains two networks: embedder and detector. Different and varying weights are associated with these two networks in the training phase to ensure imperceptibility and detectability of the watermark. To ensure robustness to attacks, we have added a set of layers with attacks between embedder and detector network. These layers are not trainable, but are required in order to determine the corresponding gradient function for propagation within the network. Attack layers have free setup parameters which are set manually. Procedure for training with attack layers should be carefully designed to preserve convergence of network and other desirable network features. This is especially important for desynchronization attacks that are particularly difficult to be handled with DNNs. Namely, these attacks tend to change the number of available samples while the DNNs work the best with fixed-size input. We are in the process of overcoming all of these issues in our research giving excellent levels of imperceptibility and robustness to all major classes of attacks. Note that our technique is developed for signals mapped to the time-frequency domain, which is the difference from classical DNN-based approaches championed by processing raw data.
Igor Djurović was born in Montenegro in 1971. He received the B.S., M.S., and Ph.D. degrees, all in electrical engineering, from the University of Montenegro, in 1994, 1996, and 2000, respectively. He is currently a Professor with the University of Montenegro and Visiting Scholar at the Lodz University of Technology. During 2002, he was on leave with the Department of Mechanical and System Engineering, Kyoto Institute of Technology, Japan, as the Japan Society for the Promotion of Science fellow. He also was a visiting professor and joint researcher with ENSIETA, Brest, France, during 2003, and visiting researcher at the Gipsa Lab, INP Grenoble, France 2009, supported by the CNRS. He was on short stays in the AIIA Laboratory, Aristotle University, Thessaloniki, the Signal Theory Group, Ruhr University, Bochum, Germany, International Center for Signal Processing, Tampere University of Technology, Tampere, Finland, National Aerospace University, Kharkov, Ukraine, etc.. He published more than 250 papers in international scientific journals and conferences. He has published eight book chapters including contribution to Time-Frequency Signal Analysis and Processing, B. Boashash, ed. He is coauthor of Time-frequency based feature extraction and classification: Considering energy concentration as a feature using the Stockwell transform and related approaches, VDM Verlag. He was member of editorial board of the Research Letters in Signal Processing, the Elsevier Signal Processing and Journal of Electrical and Computer Engineering and leading guest editor of the special issue of Eurasip Journal on Advances in Signal Processing, Special issue on Robust Signal Processing of Nonstationary Signals. He is member of the Montenegrin Academy of Sciences and Arts. He has run numerous national, regional, bilateral and international projects. He was director of the first Montenegrin Centre of Excellence in Bio-informatics (BIO-ICT). His current research interests include application of virtual instruments, time-frequency analysis-based methods for signal estimation and filtering, fractional Fourier transform applications, image processing, robust estimation, parametric and nonparametric estimation, spectral analysis, motion estimation, and digital watermarking.
Vladimir M. Vishnevsky
V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences
Experience in research and development of a vehicle identification system using RFID and broadband wireless networks
Abstract: Radio Frequency Identification (RFID) technology finds its application in many areas, including vehicle identification, logistics, production control, access control, payment services. One of the promising areas of RFID use is vehicle identification. Currently, CCTV cameras are used to solve this problem. Systems for automatic registration of traffic violations make it possible to reduce the number of accidents in which more than 15,000 people die every year in Russia, and huge material damage is caused. However, the cameras' effectiveness can be reduced to 50% or less in bad weather conditions. An effective way to eliminate this shortcoming is to equip car license plates with RFID tags and receive identification numbers over the air using RFID readers placed alongside the cameras. As the experiments show, this method of vehicle identification can provide an efficiency of up to 95% even in bad weather conditions. The timeliness and relevance of creating automated safety systems on roads using RFID technology are emphasized in the decree of the Russian Federation Government on the creation of pilot zones of such systems in Moscow, St. Petersburg, and Kazan in 2022-2024.
Developing and implementing an RFID system for vehicle identification requires solving many new complex scientific and engineering problems. This report presents an overview of the results obtained by the authors over several years of research and implementation of RFID systems for vehicles identification. We present the original results of analytical and simulation modeling of the system, showing the dependence of the probability of successful identification on the speed of vehicles with the RFID tags mounted on license plates. The models consider various settings of RFID readers, tag parameters, multipath propagation of radio signals, and the Doppler effect. The report also presents an overview of the distributed software system architecture, developed under the guidance and participation of the report's authors. This system was used in pilot implementations of the system.
A large-scale experiment in Kazan, where about 1000 vehicles were equipped with RFID tags, showed that successful identification probability reached 95%. In the pilot implementations of the system in 2020-2021 in Kazan and on the Central Ring Road in Moscow Region, readers developed at the ICS RAS and vehicles license plates with RFID tags produced by PJSC «Mikron» were used.
At the end of the report, we discuss the prospects for further development and implementation of RFID systems for vehicles identification and highlight the key problems and possible solutions.Vladimir M. Vishnevsky received the Engineering from the Moscow Institute of Electronics and Mathematics, Russia, in 1971. In 1974 and 1988, respectively, he received the Ph.D. degree in queuing theory and telecommunication networks and the D.Sc. degree in telecommunication networks. He became a Full Professor with V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences (ICS RAS) in 1989 and the Moscow Institute of Physics and Technology in 1990.
He was Deputy Director of the Institute of Information Transmission Problems of RAS from 1990 to 2010 and an Assistant Head of laboratory with ICS RAS from 1971 to 1990. He is currently the Head of Telecommunication Networks Laboratory, ICS RAS, Moscow, Russia. He is Chair of the Communications Society Chapter, IEEE Russia Section. He was a member of Expert Councils of Russian High Certifying Commission and Russian Foundation for Basic Research. He is a full member of IEEE Communication Society, International Telecommunications Academy and New York Academy of Science.
Vladimir Vishnevsky has authored over 300 papers in queuing theory and telecommunications, 10 monographs and 20 patents for inventions. His research interests lie in the areas of computer systems and networks, queuing systems, telecommunications, discrete mathematics (extremal graph theory, mathematical programming) and wireless data transmission networks.
He is a co-chair of IEEE conferences – International Congress on Ultra-Modern Telecommunications and Control Systems (ICUMT), Advances in Wireless and Optical Communications (RTUWO), and the General Chair of the International Conference on Distributed Computer and Communication Networks (DCCN). Vladimir M. Vishnevsky is chief editor of the proceedings volumes published by the Springer Publishing Company in Lecture Notes in Computer Science and Communications in Computer and Information Science series. He is Guest Editor of the special issue of the "Sensors" journal and the special issue of the “Mathematics” journal. He is an editorial board member of the following peer-reviewed journals: Automation and Remote Control, Problems of Informatics, Electronics and associate Editor-in-Chief of the scientific journal “Advances in Systems Science and Applications”.
Vladimir M. Vishnevsky is a project leader of several international research projects related to the research and development of the next generation 5G/6G networks. In 2019, by a decree of the President of the Russian Federation, Vladimir M. Vishnevsky was awarded the title "Honored Scientist of the Russian Federation".
Elena Yanakova
ELVEES RnD Center, JSC
Russian hardware and software platform ELVEES ML Platform
Abstract: the creation of a Russian software and hardware platform for managing the life cycle of machine learning from simple experiments and experiments with deep learning frameworks to training on container distributed clusters, including calculations for deploying models in an environment on a Russian ECB, is an actual task. Solving this problem will lead to the possibility of switching to a trusted Russian hardware platform in order to ensure technological security and promote the development of the Russian technological base for artificial intelligence systems.
As part of the work on creating systems with artificial intelligence, ELVIS is developing the ELVEES ML Platform software and hardware platform, which is designed to solve the problems of the life cycle of machine learning systems, combining their development, storage, deployment and operational support. The result of the work should be the creation of an MLOps infrastructure on Russian processors, which makes it possible to combine technologies and processes of machine learning and approaches to the implementation of the developed models in business processes.
The presentation provides a hardware overview of the Platform, as well as machine learning frameworks and a software stack that offer assistance in solving MLOps tasks. Currently, the Platform consists of three hardware and software systems designed to solve storage problems, train neural network algorithms and deploy on target systems, as well as two frameworks: Multicore Deep Leaning Inference Framework and Multicore Deep Leaning Training Framework, designed to solve problems neural network algorithms training
Elena Yanakova in 2006 she defended her master's thesis in the field of "Informatics and Computer Engineering" at the National Research University MIET, in 2009 she defended her Ph.D. thesis in the specialty 05.12.14 - "Radar and radio navigation". In 2014 she defended her doctoral thesis in the specialty 05.13.01 - "System Analysis, Management and Information Processing". Currently, she is the Head of the Server Solutions Laboratory of the ELVEES RnD Center, Doctor of Technical Sciences, Professor at the Institute of System and Software Engineering and Information Technologies. She is the author of more than 90 publications, including in journals included in the list of the Higher Attestation Commission, participating in Russian and international conferences, she has a number of patents for inventions.
The list of plenary papers will be expanded further...