Articles (KIV)

Permanent URI for this collection

http://hdl.handle.net/11025/30005

Browse

Showing 1 - 20 out of 195 results

Near-lossless EEG signal compression using a convolutional autoencoder: Case study for 256-channel binocular rivalry dataset
(2025) Kukrál, Martin; Pham, Duc Thien; Kohout, Josef; Kohek, Štefan; Havlík, Marek; Grygarová, Dominika
Electroencephalography (EEG) experiments typically generate vast amounts of data due to the high sampling rates and the use of multiple electrodes to capture brain activity. Consequently, storing and transmitting these large datasets is challenging, necessitating the creation of specialized compression techniques tailored to this data type. This study proposes one such method, which at its core uses an artificial neural network (specifically a convolutional autoencoder) to learn the latent representations of modelled EEG signals to perform lossy compression, which gets further improved with lossless corrections based on the user-defined threshold for the maximum tolerable amplitude loss, resulting in a flexible near-lossless compression scheme. To test the viability of our approach, a case study was performed on the 256-channel binocular rivalry dataset, which also describes mostly data-specific statistical analyses and preprocessing steps. Compression results, evaluation metrics, and comparisons with baseline general compression methods suggest that the proposed method can achieve substantial compression results and speed, making it one of the potential research topics for follow-up studies.
An algorithm for voxelised solids representation using chain codes
(2025) Repnik, Blaž; Váša, Libor; Žalik, Borut
The paper introduces a new method to describe the surfaces of voxelised solids. It operates in three stages: a hierarchical linked list of chain code sequences is created first; the linked lists are pruned; and, finally, the content of the data structure is stored. The method uses chain codes from either a three- or nine-symbols alphabet. In the first case, two chain code symbols are needed to access the next face, while, in the second case, this is done by one symbol. The pair of chain codes from the three-symbols alphabet, or the individual symbol from the nine-symbols alphabet are considered as tokens. The sets of tokens are, in both cases, extended by two tokens, indicating the beginning and ending of the list. The method processes solids of any shape, including those containing holes, cavities, or multiple components existing in the same voxel space. Edge-connectivity is permitted. The method was compared against the method proposed by Lemus et al., which is designed for solids without holes. Although supporting a broader set of voxelised solids, the proposed method generates sequences of tokens that are, on average, up to 10% shorter. Since the information entropy of the sequences of tokens produced by the proposed method is also smaller, the obtained sequences are more compressible, as confirmed by applying gzip and bzip2 data compressors.
Abilities of Contrastive Soft Prompting for Open Domain Rhetorical Question Detection
(2025) Baloun, Josef; Martínek, Jiří; Cerisara, Christophe; Král, Pavel
In this work, we start by demonstrating experimentally thatrhetorical question detection is still a challenging task, even for state-of-the-art Large Language Models (LLMs).We then propose an approach that boosts the performances of such LLMs by training a soft prompt in a waythat enables building a joint embedding space from multiple loosely related corpora.The advantages of using a soft-prompt compared to finetuning is to limit the training costs and combat overfittingand forgetting. Soft prompting is often viewed as a way to guide the model towards a specific known task, or tointroduce new knowledge into the model through in-context learning.We further show that soft prompting may also be used to modify the geometry of the embedding space, so thatthe distance between embeddings becomes semantically relevant for a target task, similarly to what is commonlyachieved with contrastive finetuning.We exploit this property to combat data scarcity for the task of rhetorical question detection bymerging several datasets into a joint semantic embedding space.We finally show on the standard Switchboard dataset that the resulting BERT-based model nearly divides by 2the number of errors as compared to Flan-T5-XXL with only 5 few-shot labeled samples, thanks to this jointembedding space. We have chosen in our experiments a BERT model because it has already been shown with S-BERT thatcontrastive finetuning of BERT leads to semantically meaningful representations. Therefore, we also show that thisproperty of BERT nicely transfers to the soft-prompting paradigm.Finally, we qualitatively analyze the resulting embedding space and propose a few heuristic criteria to selectappropriate related tasks for inclusion into the pool of training datasets.
On self-supervision in historical handwritten document segmentation
(2025) Baloun, Josef; Prantl, Martin; Lenc, Ladislav; Martínek, Jiří; Král, Pavel
Historical document analysis plays a crucial role in understanding and preserving our past. However, this task is oftenhindered by challenges such as limited annotated training data and the diverse nature of historical handwritten documents. Inthis paper,we explore the potential of self-supervised learning (SSL) in historical document analysis,with a particular focus onhistorical handwritten document segmentation, to overcome the need for extensive annotated data while enhancing efficiencyand robustness. We present an overview of SSL methods suitable for historical document analysis and discuss their potentialapplications and benefits. Furthermore, we present an approach for SSL in the document domain, considering various setups,augmentations, and resolutions. We also provide experimental results that demonstrate its feasibility and effectiveness. Ourfindings indicate that most document segmentation tasks can be effectively addressed using SSL features, highlighting thepotential of SSL to advance historical document analysis and pave the way for more efficient and robust document processingworkflows.
Classification of EEG Signal Using Deep Learning Architectures Based Motor-Imagery for an Upper-Limb Rehabilitation Exoskeleton
(2025) Khoshkhooy Titkanlou, Maryam; Pham, Duc Thien; Mouček, Roman
The brain-computer interface (BCI) is an emerging technology that enables people with physical disabilities to control and interact with devices only by using their minds and without being dependent on healthy people. One of the most popular BCI paradigms, motor imagery (MI) based on electroencephalograms (EEGs), is applied in healthcare, including rehabilitation. A significant challenge in classifying EEG signals using deep learning methods is the accurate recognition of MI signals. CNN-LSTM and CNN-Transformer are two classification algorithms proposed to improve the classification accuracy of Motor Imagery EEG signals in a noninvasive brain-computer interface. Three different methods, including noise injection (NI), conditional variational autoencoder (cVAE), and conditional GAN with Wasserstein price function and gradient penalty (cWGAN-GP), have also been implemented to augment this dataset. The best accuracy was achieved by the CNN-LSTM model, which is 79.06%, using an MI dataset involving hand movements. The dataset included 29 healthy subjects, with males aged 21–26 and females aged 18–23.
A hybrid Spiking Neural Network-Transformer architecture for motor imagery and sleep apnea detection
(2025) Pham, Duc Thien; Khoshkhooy Titkanlou, Maryam; Mouček, Roman
Introduction: Motor imagery (MI) classification and sleep apnea (SA) detection are two critical tasks in brain-computer interface (BCI) and biomedical signal analysis. Traditional deep learning models have shown promise in these domains, but often struggle with temporal sparsity and energy efficiency, especially in real-time or embedded applications.Methods: In this study, we propose SpiTranNet, a novel architecture that deeply integrates Spiking Neural Networks (SNNs) with Transformers through Spiking Multi-Head Attention (SMHA), where spiking neurons replace standard activation functions within the attention mechanism. This integration enables biologically plausible temporal processing and energy-efficient computations while maintaining global contextual modeling capabilities. The model is evaluated across three physiological datasets, including one electroencephalography (EEG) dataset for MI classification and two electrocardiography (ECG) datasets for SA detection.Results: Experimental results demonstrate that the hybrid SNN-Transformer model achieves competitive accuracy compared to conventional machine learning and deep learning models.Discussion: This work highlights the potential of neuromorphic-inspired architectures for robust and efficient biomedical signal processing across diverse physiological tasks.
Czech news dataset for semantic textual similarity
(2025) Sido, Jakub; Seják, Michal; Pražák, Ondřej; Konopík, Miloslav; Moravec, Václav
This paper describes a novel dataset consisting of sentences with two different semantic similarity annotations; with and without surrounding context. The data originate from the journalistic domain in the Czech language. The final dataset contains 138,556 human annotations divided into train and test sets. In total, 485 journalism students participated in the creation process. To increase the reliability of the test set, we compute the final annotations as an average of 9 individual annotation scores. We evaluate the dataset quality by measuring inter and intra-annotator agreements. Besides agreement numbers, we provide detailed statistics of the collected dataset. We conclude our paper with a baseline experiment of building a system for predicting the semantic similarity of sentences. Due to the massive number of training annotations (116,956), the model significantly outperforms an average annotator (0.92 versus 0.86 of Pearson’s correlation coefficient).
Czech medical coding assistant based on transformer networks
(2024) Lenc, Ladislav; Martínek, Jiří; Baloun, Josef; Přibáň, Pavel; Prantl, Martin; Taylor, Stephen; Král, Pavel; Kyliš, Jiří
Clinical coding of medical reports is currently carried out manually by a so-called clinical coder. However, due to the human factor, this process is error-prone and expensive. The coder needs to be properly trained and spends significant effort on each report, leading to occasional mistakes. The main goal of this paper is to propose and implement a system that serves as an assistant to the coder and automatically predicts diagnosis codes. These predictions are then presented to the coder for approval or correction, aiming to enhance efficiency and accuracy. The main contribution lies in the proposal and evaluation of ICD classification models for the Czech language with relatively few training parameters, allowing swift utilisation on the prevalent computer systems within Czech hospitals and enabling easy retraining or fine-tuning with newly available data. First, we introduce a small transformer-based model for each task followed by the design of a transformer-based “Four-headed” model incorporating four distinct classification heads. This model achieves comparable, sometimes even better results, against four individual models. Moreover this novel model significantly economises memory usage and learning time. We also show that our models achieve comparable results against state-of-the-art English models on the Mimic IV dataset even though our models are significantly smaller.
Efficient sleep apnea detection using single-lead ECG: A CNN-Transformer-LSTM approach
(2025) Pham, Duc Thien; Mouček, Roman
Background:Sleep apnea (SA), a prevalent sleep-related breathing disorder, disrupts normal respiratory patterns during sleep. This disruption can have a cascading effect on the body, potentially leading to complications in various organs, including the heart, brain, and lungs. Due to the potential for these complications, early and accurate detection of SA is critical. Electrocardiograms (ECG), due to their ability to continuously monitor heart rhythms and detect subtle changes in cardiac activity, such as heart rate variability and arrhythmias, which are often linked to sleep disruptions, have become crucial in identifying individuals at risk for SA.Method:In this study, we propose a hybrid neural network model named CNN-Transformer-LSTM that uses a single-lead ECG signal to detect SA automatically. This method captures spatial and temporal features in the ECG data to improve classification performance. Our model utilizes RR intervals (RRI) and R-peak signals derived from ECG data as input and then classifies SA and normal states on a per-segment and per-recording basis. We evaluated the model using the Physionet Apnea-ECG dataset, consisting of 70 single-lead ECG recordings annotated by medical professionals, and the UCD St. Vincent’s University Hospital’s sleep apnea database (UCDDB) containing polysomnogram records from 25 patients.Results:Our model achieved an accuracy of 91.6% for per-segment classification on the Physionet Apnea-ECG dataset using hold-out validation and the highest accuracy of 94.1% using five-fold cross-validation. As for per-recording classification, our model achieved an accuracy of 100% and the highest correlation coefficient value of 0.9996 using five-fold cross-validation. On the UCDDB dataset, our model achieved an accuracy of 99.37% on the reduced dataset excluding 4 patients and 98.34% on the full dataset. Compared to previous works, our model improved the per-segment classification accuracy by nearly 3% over the existing best result, thereby demonstrating that our model outperforms existing state-of-the-art methods in accurately detecting SA from a single-lead ECG signal.Conclusion:These results highlight the effectiveness of the CNN-Transformer-LSTM model for SA detection and its potential to be used in SA detection devices for home health care and clinical settings.
Simple Derivation of the Hermite Bicubic Patch using Tensor product
(2025) Skala, Václav
Bicubic parametric patches are widely used in various geometric applications. These patchesare critical in CAD/CAM systems, which are applied in the automotive industry and mechanical andcivil engineering. Commonly, Hermite, B ́ezier, Coons, or NURBS patches are employed in practice.However, the construction of the Hermite bicubic patch is often not easy to explain formally. Thiscontribution presents a new formal method for constructing the Hermite bicubic plate based on thetensor product approach
An Efficient point-in-convex 3D polyhedron test using a projective algorithm with sub-linear expected complexity
(2025) Skala, Václav
We propose a novel algorithm for determining whether a given point lies within a convex polyhedron, achieving a sub-linear computational expected complexity of Oexp(N1/2), where N represents the number of triangles in the polyhedron’s triangular mesh. In contrast to traditional methods with linear complexity O(N), our approach significantly reduces com-putational overhead, making it especially effective for large polyhedral models. The algorithm is formulated entirely in projective space, utilizing homogeneous coordinates for the tested points and triangle vertices. By leveraging vector–vec¬tor operations optimized for SSE, AVX instructions, and GPU architectures, our method is robust and straightforward, tailored to handle even highly complex convex polyhedra. The efficiency of the approach was validated through theoretical analysis and estimated speed-up calculations, demonstrating its potential to accelerate applications in computer graphics, computational geometry, collision detection, and related fields. Additionally, the simplicity of the proposed algorithm ensures a high potential for broad applicability and supports further advancements in this area.
Register-Based and Stack-Based Virtual Machines: Which Perform Better in JIT Compilation Scenarios?
(2025) Šimek, Bohuslav; Fiala, Dalibor; Dostal, Martin
Background: Just-In-Time (JIT) compilation plays a critical role in optimizing the performance of modern virtual machines (VMs). While the architecture of VMs – register-based or stack-based – has long been a subject of debate, empirical analysis focusing on JIT compilation performance is relatively sparse. Objective: In this study, we aim to answer the question: “Register-based and stack-based virtual machines: which perform better in JIT compilation scenarios?”. Methods: We explore this through a comprehensive set of benchmarks measuring execution speed. To achieve this, we developed identical test cases in languages that support both types of VM architectures and ran these tests under controlled conditions. The performance metrics were captured and analyzed for JIT compilation, including initial interpretation, bytecode translation, and optimized code execution. Results: Our findings suggest that register-based VMs generally outperform stack-based VMs in terms of execution speed. Moreover, the performance gap between the two architectures in mixed execution mode, which essentially copies characteristics of the underlying virtual machine, suggests that making the right choice of VM architecture is still important. Conclusion: This study provides developers, researchers, and system architects with actionable insights into the performance trade-offs associated with each VM architecture in JIT-compiled environments. The findings can guide the design decisions in the development of new virtual machines and JIT compilation strategies.
Robust Line-Convex Polygon Intersection Computation in E2 using Projective Space Representation
(2023) Skala, Václav
This paper describes modified robust algorithms for a line clipping by a convex polygon inE2and a convex polyhedron inE3. The proposed algorithm is based on the Cyrus-Beck algorithmand uses homogeneous coordinates to increase the robustness of computation. The algorithm enablescomputation fully in the projective space using the homogeneous coordinates and the line can be givenin the projective space, in general. If the result can remain in projective space, no division operation isneeded. It supports the use of vector-vector operations, SSE/AVX instructions, and GPU.
A New Fully Projective O(lg N) Line Convex Polygon Intersection Algorithm
(2025) Skala, Václav
Intersecting algorithms, especially line clipping in E2 and E3 in computer graphics, have been studied for a long time. Many different algorithms have been developed. The simplest case is a line clipping by a convex polygon in E2 with O(N) computational complexity and with known polygon edges orientation. This contribution presents a new algorithm for a line clipping by a convex polygon in E2 with O(lg N) complexity, which is based on the point-in-half plane test. The proposed algorithm does not require prior knowledge of the polygon edge orientation. The vertices of the convex polygon and the clipped line can be given in projective space using homogeneous coordinates. The algorithm uses vector–vector operations for efficient implementation with SSE or AVX vector–vector instructions or on GPUs.
A new fully projective O(log N) point-in-convex polygon algorithm: a new strategy
(2025) Skala, Václav
A novel and fully projective algorithm for a point-in-convex polygon test with computational complexity of O(log N) in 2D isdescribed in this contribution. The polygon vertices and tested points can be given in projective space without conversion toEuclidean space. The proposed algorithm is simple, robust, easy to implement, and invariant to the convex polygon orientation.It can be easily modified for use in Euclidean space and CPU implementation. Vector–vector operations are used, makingit suitable for implementation using SSE, AVX, and FMA instructions.
On self-supervision in historical handwritten document segmentation
(Springer, 2025) Baloun, Josef; Prantl, Martin; Lenc, Ladislav; Martínek, Jiří; Král, Pavel
Russian Publications in Web of Science: A Bibliometric Study
(2023) Fiala, Dalibor; Maltseva, Daria
This article presents a bibliometric study of 1.38 million Russian publications indexed in Web of Science as of May 2022 without any restrictions as to document types, time periods, scientific disciplines, etc. From this perspective, the present analysis reflects Russian research’s true presence and visibility in the most prestigious scientific literature database. The main results obtained are: a) There was a rapid increase in research production in the 2010s, but the share of the Russian output in the global research production is still below 3%. b) International collaborative publications account for about 30% of Russian papers but around 70% of Russian citations. c) Physics, chemistry, and engineering are the most productive Russian research areas, but their citation impact is below the world average in those respective fields. d) The most frequently collaborating countries are the United States, Germany, and France, but Canada and Switzerland consistently contribute to the greatest relative citation impact of collaborative papers in the top ten research areas.
Analysis of Cited References in Russian Publications on Web of Science
(2024) Fiala, Dalibor; Maltseva, Daria
In this article we analyze the cited references in 1.38 million papers by Russian (co-)authors indexed in the Web of Science database until May 2022. Similarly, to the established processes in the so-called Reference Publication Year Spectroscopy (RPYS), we study the distribution of the references across the cited years and seek to identify the peak years with the publications that attracted the most attention of Russian scholars. In this way, the historical roots of Russian science may be traced and we take a closer look at these most influential works. In addition, we investigate the evolution of the mean age of references and of their average number per paper over time and inspect the most frequently cited sources. The results show that the average number of references in Russian papers has been steadily increasing, but the mean age of references has been declining in the most recent years. Also, the foundations of Russian science seem to be physics of particles and electrochemistry and have recently become based more internationally than in the past. This study is the first of its kind and may help better understand the character of Russian research.
Editing mesh sequences with varying connectivity
(2024) Hácha, Filip; Dvořák, Jan; Káčereková, Zuzana; Váša, Libor
Time-varying connectivity of triangle mesh sequences leads to substantial difficulties in their processing. Unlike editing sequences with constant connectivity, editing sequences with varying connectivity requires addressing the problem of temporal correspondence between the frames of the sequence. We present a method for time-consistent editing of triangle mesh sequences with varying connectivity using sparse temporal correspondence, which can be obtained using existing methods. Our method includes a deformation model based on the usage of the sparse temporal correspondence, which is suitable for the temporal propagation of user-specified deformations of the edited surface with respect to the shape and true topology of the surface while preserving the individual connectivity of each frame. Since there is no other method capable of comparable types of editing on time-varying meshes, we compare our method and the proposed deformation model with a baseline approach and demonstrate the benefits of our framework.
Accelerated multi-hillshade hierarchic clustering for automatic lineament extraction
(2024) Kaas, Ondřej; Šilhavý, Jakub; Kolingerová, Ivana; Čada, Václav
The lineaments are linear features reflecting mountain ridges or discontinuities in the geological structure. Recently, an automatic approach of their recognition based on multi-hillshade hierarchic clustering (MHHC) has been developed, based on line extraction from a raster image. This paper presents a modification of MHHC, which solves the spatial line segment clustering as a facility location problem. The proposed modification is faster than MHHC while not changing the method’s core.

Browse

Recent Submissions