Tomasz P. Stefanski
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Tomasz P. Stefanski has a total of 40 co-authors. Here are the top co-authors by number of shared publications:
- Jacek Gulgowski
Faculty of Mathematics, Physics and Informatics, University of Gdańsk, Gdańsk, Poland
12 shared publications - Kosmas L. Tsakmakidis
Department of Physics, School of Science, National and Kapodistrian University of Athens, Zografou Panepistimioupolis, 15784 Athens, Greece
6 shared publications Konstantinos Baskourelos
Solid State Physics Section, Department of Physics, National and Kapodistrian University of Athens 1 , Panepistimioupolis, GR - 157 84, Athens, Greece
3 shared publications- Łukasz Grzymkowski
Arrow Electronics, 80-309 Gdansk, Poland
3 shared publications - Anna Jezierska
Department of Modelling and Optimization of Dynamical Systems, Systems Research Institute, Warsaw, Poland
2 shared publications - Artur Noga
Department of Electronics, Electrical Engineering and Microelectronics, Silesian University of Technology, Akademicka 16, 44-100 Gliwice, Poland
2 shared publications
Latest Publications
- 6 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33 (10709908) , 1756-1760
A Lightweight RAW Video Denoiser With Coarse Alignment and Imbalanced Noise Pre-Training
Deep neural networks have significantly advanced video denoising. However, deploying accurate denoisers across platforms remains challenging, particularly under real-time constraints and with high-resolution inputs. While explicit motion estimation modules improve temporal alignment and enhance denoising quality, they are often omitted in lightweight models due to their computational burden. This paper investigates the trade-off between denoising quality and computational efficiency when using spatially coarse optical flow, and shows that reduced-resolution flow retains most temporal denoising benefits while significantly lowering computational cost. In addition, we introduce an inference-free imbalanced noise pre-training strategy that encourages effective temporal aggregation and complements coarse alignment without modifying the runtime architecture. The proposed Lightweight Video Denoiser (LVD) achieves state-of-the-art performance among lightweight models on RAW video denoising benchmarks while maintaining low latency suitable for real-time edge deployment. Our code is available at https://github.com/camalab-ai/LVD
- 6 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33 (10709908) , 1756-1760
A Lightweight RAW Video Denoiser With Coarse Alignment and Imbalanced Noise Pre-Training
Deep neural networks have significantly advanced video denoising. However, deploying accurate denoisers across platforms remains challenging, particularly under real-time constraints and with high-resolution inputs. While explicit motion estimation modules improve temporal alignment and enhance denoising quality, they are often omitted in lightweight models due to their computational burden. This paper investigates the trade-off between denoising quality and computational efficiency when using spatially coarse optical flow, and shows that reduced-resolution flow retains most temporal denoising benefits while significantly lowering computational cost. In addition, we introduce an inference-free imbalanced noise pre-training strategy that encourages effective temporal aggregation and complements coarse alignment without modifying the runtime architecture. The proposed Lightweight Video Denoiser (LVD) achieves state-of-the-art performance among lightweight models on RAW video denoising benchmarks while maintaining low latency suitable for real-time edge deployment. Our code is available at https://github.com/camalab-ai/LVD
- Open Access
- 3 April 2026
- Published by MDPI AG in Electronics
- Vol. 15 (7) , 1508
What Is That Noise: Survey of Anomalous Sound Detection Using Edge Systems
In this paper, we provide a thorough review of novel machine learning (ML) models for anomalous sound detection (ASD). We focus on deploying models to highly constrained, embedded systems and tiny ML, and using single-channel sound as the data input. The survey includes only the works published in 2020 and later. Researchers address the anomaly detection task in various ways, borrowing models and techniques from such fields as speech processing, audio generation, and even computer vision. However, it is not clear which of these are suitable for embedded systems, meeting their constraints such as memory or compute. To address that, we provide a deep analysis of these models and optimization techniques applied to meet the design criteria for embedded platforms. We consider both deep learning and classical ML methods. We define categories for the anomaly detection methods depending on the approach taken to provide a structure and simplify the comparison of methods. We aim to provide a guideline on how to develop ASD systems and how to efficiently deploy the models on the embedded platforms.
- 3 April 2026
- Published by MDPI AG in Electronics
- Vol. 15 (7) , 1508
Open AccessWhat Is That Noise: Survey of Anomalous Sound Detection Using Edge Systems
In this paper, we provide a thorough review of novel machine learning (ML) models for anomalous sound detection (ASD). We focus on deploying models to highly constrained, embedded systems and tiny ML, and using single-channel sound as the data input. The survey includes only the works published in 2020 and later. Researchers address the anomaly detection task in various ways, borrowing models and techniques from such fields as speech processing, audio generation, and even computer vision. However, it is not clear which of these are suitable for embedded systems, meeting their constraints such as memory or compute. To address that, we provide a deep analysis of these models and optimization techniques applied to meet the design criteria for embedded platforms. We consider both deep learning and classical ML methods. We define categories for the anomaly detection methods depending on the approach taken to provide a structure and simplify the comparison of methods. We aim to provide a guideline on how to develop ASD systems and how to efficiently deploy the models on the embedded platforms.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 120147-120163
FinetextPL-Edu: A High-Quality Polish Corpus for Large Language Model Pretraining via Educational Quality Filtering
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 120147-120163
Open AccessFinetextPL-Edu: A High-Quality Polish Corpus for Large Language Model Pretraining via Educational Quality Filtering
- Open Access
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 35 (12) , 12196-12211
Lifting Deep Image Denoisers to Video With Frame Interpolation Pre-Training
We introduce Frame Interpolation Pre-training (FIP), a simple learning technique for lifting deep image denoisers to video denoising with improved implicit temporal alignment. Modern video denoising networks typically rely on explicit motion estimation and alignment which are computationally intensive and harder to re-design and re-train, restricting their application scope and usability. Conversely, stacking frames and image denoisers, without incorporating explicit motion estimation modules, improves speed and benefits from a simpler design, thereby facilitating their generalizability to the video domain. However, it leads to lower accuracy due to suboptimal capture of temporal dependencies. To better leverage the adjacent frames in this setting and reduce the accuracy gap, we propose a novel training regime that divides the standard supervised training of the denoising task into two phases. In the initial phase, FIP guides the network to interpolate a fully masked central frame using only adjacent noisy input frames. In the subsequent phase, the pre-trained network is fine-tuned on denoising the central frame, now using all noisy input frames. Extensive diagnostics indicate that FIP-based networks provide better implicit motion estimation and temporal alignment. In effect, qualitative and quantitative evaluation on standard video denoising datasets with synthetic and real noise demonstrates that FIP consistently improves video denoising accuracy of motion-aware, video-lifted image denoisers without additional computational overhead during training and test time. Our code is available at https://github.com/camalab-ai/FIP.
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 35 (12) , 12196-12211
Open AccessLifting Deep Image Denoisers to Video With Frame Interpolation Pre-Training
We introduce Frame Interpolation Pre-training (FIP), a simple learning technique for lifting deep image denoisers to video denoising with improved implicit temporal alignment. Modern video denoising networks typically rely on explicit motion estimation and alignment which are computationally intensive and harder to re-design and re-train, restricting their application scope and usability. Conversely, stacking frames and image denoisers, without incorporating explicit motion estimation modules, improves speed and benefits from a simpler design, thereby facilitating their generalizability to the video domain. However, it leads to lower accuracy due to suboptimal capture of temporal dependencies. To better leverage the adjacent frames in this setting and reduce the accuracy gap, we propose a novel training regime that divides the standard supervised training of the denoising task into two phases. In the initial phase, FIP guides the network to interpolate a fully masked central frame using only adjacent noisy input frames. In the subsequent phase, the pre-trained network is fine-tuned on denoising the central frame, now using all noisy input frames. Extensive diagnostics indicate that FIP-based networks provide better implicit motion estimation and temporal alignment. In effect, qualitative and quantitative evaluation on standard video denoising datasets with synthetic and real noise demonstrates that FIP consistently improves video denoising accuracy of motion-aware, video-lifted image denoisers without additional computational overhead during training and test time. Our code is available at https://github.com/camalab-ai/FIP.
- Open Access
- 2 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 181041-181051
Global Roots and Poles Finding Algorithm on Quantum Computer
In this paper, the implementation of the global roots and poles finding algorithm for a complex-valued function of a complex variable on a quantum computer, which allows for solving general nonlinear algebraic equations, is presented. The considered function is sampled with the use of Delaunay’s triangulation on the complex plane and a phase quadrant, in which the value of the function is located, is computed on a classical computer for all of the sampling nodes. Then, if the real and imaginary parts of the function simultaneously change signs for both ends of the same edge in the mesh, then a zero of the function is located in the region around this edge. In order to detect such edges, the mesh is transformed into a one-dimensional array and the required edges, where the sign simultaneously changes for real and imaginary parts of the function, are found with the use of quantum Grover’s algorithm. If the mesh consists of P edges, the computational overhead of this operation, in terms of oracle queries, is equal to O (√ P ) on a quantum computer, instead of O ( P ) on a classical one. Finally, the existence of function zeros and poles is proved with the use of Cauchy’s argument principle on a classical computer, and the output results are computed, based on the mesh refinement, with the assumed numerical precision of computations. Our method is implemented in Python with the use of the Qiskit software development kit and its applicability is proved by quantum emulations.
- 2 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 181041-181051
Open AccessGlobal Roots and Poles Finding Algorithm on Quantum Computer
In this paper, the implementation of the global roots and poles finding algorithm for a complex-valued function of a complex variable on a quantum computer, which allows for solving general nonlinear algebraic equations, is presented. The considered function is sampled with the use of Delaunay’s triangulation on the complex plane and a phase quadrant, in which the value of the function is located, is computed on a classical computer for all of the sampling nodes. Then, if the real and imaginary parts of the function simultaneously change signs for both ends of the same edge in the mesh, then a zero of the function is located in the region around this edge. In order to detect such edges, the mesh is transformed into a one-dimensional array and the required edges, where the sign simultaneously changes for real and imaginary parts of the function, are found with the use of quantum Grover’s algorithm. If the mesh consists of P edges, the computational overhead of this operation, in terms of oracle queries, is equal to O (√ P ) on a quantum computer, instead of O ( P ) on a classical one. Finally, the existence of function zeros and poles is proved with the use of Cauchy’s argument principle on a classical computer, and the output results are computed, based on the mesh refinement, with the assumed numerical precision of computations. Our method is implemented in Python with the use of the Qiskit software development kit and its applicability is proved by quantum emulations.
- Open Access
- 14 November 2024
- Published by MDPI AG in Electronics
- Vol. 13 (22) , 4472
Efficient Analysis of Noise Induced in Low-Voltage Installations Placed Inside Buildings with Lightning Protection Systems
This paper describes an efficient approach to the broadband analysis of lightning protection systems (LPSs) using the method of moments (MoM) implemented in the frequency domain. The adaptive frequency sampling (AFS) algorithm, based on a rational interpolation of the relevant observable (e.g., voltage, current, electric or magnetic field) which describes the properties of the LPS, is employed to reduce the number of samples computed by the full-wave MoM. This improvement is achieved by the quick comparison of two interpolants with the use of the Stöer–Bulirsch algorithm, which provides the frequency location of the next MoM samples for computations. This algorithm allows for the efficient localization of resonant frequencies while reducing the number of samples computed over the entire frequency range. In the instances when the induced noise is determined in low-voltage installations protected by various types of LPSs, reductions in computational overhead equal to 47.9× and 72.1× in broadband LPS simulations are obtained. Hence, the proposed approach allows for a significant reduction in computational overhead in comparison to standard, uniformly sampled simulations.
- 14 November 2024
- Published by MDPI AG in Electronics
- Vol. 13 (22) , 4472
Open AccessEfficient Analysis of Noise Induced in Low-Voltage Installations Placed Inside Buildings with Lightning Protection Systems
This paper describes an efficient approach to the broadband analysis of lightning protection systems (LPSs) using the method of moments (MoM) implemented in the frequency domain. The adaptive frequency sampling (AFS) algorithm, based on a rational interpolation of the relevant observable (e.g., voltage, current, electric or magnetic field) which describes the properties of the LPS, is employed to reduce the number of samples computed by the full-wave MoM. This improvement is achieved by the quick comparison of two interpolants with the use of the Stöer–Bulirsch algorithm, which provides the frequency location of the next MoM samples for computations. This algorithm allows for the efficient localization of resonant frequencies while reducing the number of samples computed over the entire frequency range. In the instances when the induced noise is determined in low-voltage installations protected by various types of LPSs, reductions in computational overhead equal to 47.9× and 72.1× in broadband LPS simulations are obtained. Hence, the proposed approach allows for a significant reduction in computational overhead in comparison to standard, uniformly sampled simulations.
- 1 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Antennas and Wireless Propagation Letters
- Vol. 23 (8) , 2446-2450
Absorbing Boundary Conditions Derived Based on Pauli Matrices Algebra
In this letter, we demonstrate that a set of absorbing boundary conditions (ABCs) for numerical simulations of waves, proposed originally by Engquist and Majda and later generalized by Trefethen and Halpern, can alternatively be derived with the use of Pauli matrices algebra. Hence a novel approach to the derivation of one-way wave equations in electromagnetics is proposed. That is, the classical wave equation can be factorized into two two-dimensional wave equations with first-order time derivatives. Then, using suitable approximations, not only Engquist and Majda ABCs can be obtained, but also generalized ABCs proposed by Trefethen and Halpern, which are applicable to simulations of radiation problems.
- 1 May 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Antennas and Wireless Propagation Letters
- Vol. 23 (8) , 2446-2450
Absorbing Boundary Conditions Derived Based on Pauli Matrices Algebra
In this letter, we demonstrate that a set of absorbing boundary conditions (ABCs) for numerical simulations of waves, proposed originally by Engquist and Majda and later generalized by Trefethen and Halpern, can alternatively be derived with the use of Pauli matrices algebra. Hence a novel approach to the derivation of one-way wave equations in electromagnetics is proposed. That is, the classical wave equation can be factorized into two two-dimensional wave equations with first-order time derivatives. Then, using suitable approximations, not only Engquist and Majda ABCs can be obtained, but also generalized ABCs proposed by Trefethen and Halpern, which are applicable to simulations of radiation problems.
- 22 November 2023
- Published by Elsevier in Communications in Nonlinear Science and Numerical Simulation
- Vol. 129
Modelling and simulations in time-fractional electrodynamics based on control engineering methods
- 22 November 2023
- Published by Elsevier in Communications in Nonlinear Science and Numerical Simulation
- Vol. 129
Modelling and simulations in time-fractional electrodynamics based on control engineering methods
- Open Access
- 25 September 2023
- Published by MDPI AG in Applied Sciences
- Vol. 13 (19) , 10654
FDTD Method for Electromagnetic Simulations in Media Described by Time-Fractional Constitutive Relations
In this paper, the finite-difference time-domain (FDTD) method is derived for electromagnetic simulations in media described by the time-fractional (TF) constitutive relations. TF Maxwell’s equations are derived based on these constitutive relations and the Grünwald–Letnikov definition of a fractional derivative. Then the FDTD algorithm, which includes memory effects and energy dissipation of the considered media, is introduced. Finally, one-dimensional signal propagation in such electromagnetic media is considered. The proposed FDTD method is derived based on a discrete approximation of the Grünwald–Letnikov definition of the fractional derivative and evaluated in a code. The stability condition is derived for the proposed FDTD method based on a numerical-dispersion relation. The obtained numerical results are compared with the outcomes of reference frequency-domain simulations, proving the accuracy of the proposed approach. However, high spatial resolution is required in order to obtain accurate results. The developed FDTD method is, unfortunately, computation and memory demanding when compared to the ordinary FDTD algorithm.
- 25 September 2023
- Published by MDPI AG in Applied Sciences
- Vol. 13 (19) , 10654
Open AccessFDTD Method for Electromagnetic Simulations in Media Described by Time-Fractional Constitutive Relations
In this paper, the finite-difference time-domain (FDTD) method is derived for electromagnetic simulations in media described by the time-fractional (TF) constitutive relations. TF Maxwell’s equations are derived based on these constitutive relations and the Grünwald–Letnikov definition of a fractional derivative. Then the FDTD algorithm, which includes memory effects and energy dissipation of the considered media, is introduced. Finally, one-dimensional signal propagation in such electromagnetic media is considered. The proposed FDTD method is derived based on a discrete approximation of the Grünwald–Letnikov definition of the fractional derivative and evaluated in a code. The stability condition is derived for the proposed FDTD method based on a numerical-dispersion relation. The obtained numerical results are compared with the outcomes of reference frequency-domain simulations, proving the accuracy of the proposed approach. However, high spatial resolution is required in order to obtain accurate results. The developed FDTD method is, unfortunately, computation and memory demanding when compared to the ordinary FDTD algorithm.