The primary objective of this work was to automate the transfer of digitized invoices into an accounting software system. Traditionally, manual invoice processing is time-consuming, prone to errors, and increases operational costs, hindering timely financial reporting. The work analyzes various solutions for invoice digitization, their export methods, and the technologies utilized for system development and data transmission. The key outcome is the successful implementation of a system for transferring digitized invoices into accounting software. The system leverages a modular design with a microservices-oriented architectural approach, with C# chosen as the primary programming language. To ensure reliable asynchronous communication, AWS Simple Queue Service (SQS) was utilized, and AWS DynamoDB was used for persistent data storage. The system was designed to eliminate manual intervention in the invoice processing workflow, thereby enhancing efficiency, accuracy, and overall operational speed. This work significantly contributes to centralizing invoices in a digital format with a unified design, streamlining accounting processes, and laying the groundwork for advanced financial analytics or machine learning applications in accounting.
The volatility of cryptocurrencies poses challenges for accurate price forecasting. This study compares traditional machine learning models (Linear Regression, Support Vector Regression), ensemble methods (Random Forest, Gradient Boosting), and deep learning architectures (LSTM, GRU) in predicting the daily prices of Bitcoin (BTC), Ethereum (ETH), Cardano (ADA), Ripple (XRP), and Polkadot (DOT). Using historical data from CoinGecko, a sliding window, and normalisation, we assess models by Root Mean Square Error (RMSE) and relative percentage error. Results show that LSTM and GRU achieve the best overall accuracy, while Linear Regression remains competitive for stable assets such as BTC, ADA, and DOT. Ensemble methods performed moderately, whereas SVR consistently underperformed. The findings underline the importance of matching prediction models to the characteristics of specific cryptocurrencies.
This paper presents a cloud-based facial biometric authentication system integrating demographic and emotional attributes to enhance verification robustness under compression and distortion. A prototype web platform was implemented using cloud-based services, combining facial recognition with demographic and emotional analysis. The system was experimentally evaluated under multiple image degradation scenarios, including lossy and lossless compression, blurring, and block partitioning. Results show that recognition accuracy remains above 98% under mild compression, but decreases significantly under severe distortions, with match rates dropping below 25% in extreme cases. Additional experiments demonstrate reliable gender classification (95%) and moderate performance for age and emotion estimation. The contribution of this work lies in demonstrating the practical feasibility of combining identity verification with contextual biometric features in resource-constrained environments, while also outlining the limitations and security implications of such systems.
This article presents the design and validation of a virtual protection system concept utilizing a Python-based software environment. Unlike traditional hardware-centric relays, this approach leverages software-defined logic to enhance flexibility and scalability in distribution networks. The study focuses on the development of a virtual protection architecture, including a custom-built input signal generator to simulate fault conditions without physical hardware. Specific protection functions, including ANSI-59PGQ (Overvoltage) and ANSI-27 (Undervoltage), were implemented and tested. The system demonstrated a processing latency of 5–20 ms during simulation cycles, validating the algorithmic efficiency. While the current prototype relies on soft real-time simulation, the results confirm that the proposed virtual protection logic is capable of detecting faults with reaction times comparable to commercial digital relays, paving the way for future hardware-integrated implementations.
This study explores the application of optical flow techniques for the automated classification of lung sliding in lung ultrasound—a key indicator for assessing lung function and diagnosing pneumothorax. Optical flow is used to capture subtle pleural motion between frames. We evaluate both traditional machine learning models (Random Forest, Gradient Boosting) and deep learning architectures (CNN, ResNet-18), using various preprocessing methods to optimize flow representation. Among deep models, the CNN achieved the highest overall accuracy (80.8%) and F1-score (73.7%), while the more complex ResNet-18 reached a high recall (94.5%) but suffered from lower precision (62.6%), indicating a tendency to over-predict the positive class. These findings highlight the trade-offs between model complexity and generalization in limited-data scenarios. The results underline the importance of model selection and tailored preprocessing in improving diagnostic reliability.
This paper investigates and extends the capabilities of KANICE Mini, a hybrid neural architecture that integrates the Kolmogorov – Arnold Network framework with Interactive Convolution Elements. While the original implementation achieved 99.35% accuracy on the MNIST dataset, we improve the model’s performance through a refined training pipeline, enhanced regularization techniques, and structured hyperparameter optimization. Our optimized KANICE Mini achieves 99.56% accuracy on MNIST, surpassing the original result. Furthermore, we evaluate its generalization capability on more complex real-world data by applying it to the Invasive Ductal Carcinoma classification task, where it reaches 85.79% accuracy. These results demonstrate that, with careful tuning, KANICE Mini can rival significantly larger architectures in performancewhile preserving advantages in efficiency, modularity, and interpretability.
This paper investigates neural network models for predicting weather parameters and received signal strength indicator (RSSI) to enable adaptive handover in hybrid free space optics (FSO)/radio frequency (RF) systems. The most correlated parameters were visibility, temperature (measured from three independent sensors), and particle concentration. The work was exclusively focused on predicting correlated atmospheric effects and subsequently predicting RSSI parameter derived from them using a long short-term memory (LSTM) model, minute by minute over 24 hours. The predicted values can in the future serve as input for initiating handover decisions – for example, through visualization in Simulink. The threshold value of -30 dBm was taken from peer-reviewed articles and is used as a reference limit for link quality.
This paper presents a comprehensive comparative analysis of fuzzy logic and conventional PI controllers in motor control applications, focusing on dynamic load scenarios. We investigate the integration of fuzzy logic controllers within field-oriented control (FOC) architectures, particularly examining their performance in speed regulation, oscillation reduction, and transient response. Through systematic experimentation with a sensorless PMSM drive system, we demonstrate how fuzzy controllers can outperform traditional PI regulators in handling nonlinearities and varying operating conditions. The study includes detailed analysis of control surface optimization, revealing how strategic modifications to membership functions can significantly reduce speed oscillations while maintaining responsiveness. Our results show that the implemented fuzzy speed controller achieves 28.57% faster settling time compared to its PI counterpart, albeit with a 2% overshoot trade-off. Furthermore, we identify key subsystems within FOC architectures where fuzzy logic implementation offers the most substantial benefits, including open-loop start-up sequences and flux control. The paper concludes with practical insights into computational requirements for embedded implementations and discusses future directions for intelligent control systems in motor drive applications.
This study contributes to the development of multifunctional shielding textiles effective against both electromagnetic interference (EMI) and ionizing radiation. Wool-blend fabrics were modified with Fe₃O₄ magnetic nanoparticles to create flexible composites with strong frequency-dependent absorption and reflection of electromagnetic fields, along with protection against alpha and beta particles. Magnetic susceptibility analysis revealed that, in some cases, radiation exposure unexpectedly enhanced the material’s magnetic properties, suggesting beneficial structural changes in the nanoparticles. These findings open new possibilities for adaptive “smart” textiles that not only withstand radiation but may improve under it—paving the way for use in wearable electronics, aerospace, and next-generation protective clothing.
This article describes the expansion of an existing data processing platform. This work aimed to analyze the current state of the platform, identify possible shortcomings, and propose new containers. These containers were later implemented and connected with the rest of the platform. The extensions described in this article provide new ways to load data from different databases or video sources, new processing tasks, and an email notification system utilizing Gmail. The results of this work provide a great stepping stone in creating the fully featured platform, which we plan on creating. It also serves as a confirmation of the ease of implementing new containers into our framework.
The most used asymmetric encryption algorithm nowadays is RSA. It may become insecure regarding advances in the field of quantum computers. It is the reason why the National Institute of Standards and Technologies introduces the challenges of choosing a new post-quantum encryption standard. Initially, NIST received 82 submissions comprising key encapsulation mechanisms and digital signature schemes. However, only 69 of them were formally accepted after an initial review. In July 2022, NIST selected some algorithms for standardization. For the key encapsulation mechanism, Kyber was selected, and for digital signatures, Dilithium Falcon and SPHINCS+. After the third round concluded, NIST indicated it would continue to evaluate some of the alternative algorithms that were not selected as finalists in the third round. This ongoing evaluation is informally referred to as a "fourth round." Initially, there were four participants - BIKE, Classic McEliece, SIKE, and HQC. However, the SIKE downfall with the Castryck-Decru attack was introduced in July 2022, and the HQC algorithm was chosen for standardization in March 2025. In our research, we examine all functions of BIKE, Classic McEliece, and the HQC from the point of view of time and memory consumption. The results obtained will help us during the implementation of the BIKE algorithm on ESP32.
Dynamic gesture recognition systems face persistent challenges in achieving real-time performance and high recognition efficiency. This paper presents a novel framework integrating computer vision techniques with machine learning algorithms to address these issues. The approach utilizes OpenCV for dynamic gesture detection, employing gesture contour extraction using skin color feature segmentation, fingertip detection for static gesture recognition, and dynamic gesture tracking through Hidden Markov Models (HMM). Experimental results demonstrate high recognition accuracy of 95.8 ± 1.4% across various gesture types, with individual rates ranging from 95.2 ± 1.6% to 98.5 ± 0.8%. Real-time performance is achieved with average processing time of 82.71 ± 3.2 ms per frame (12.1 FPS). Our method demonstrates 1.15× to 1.49× speed improvements over state-of-the-art approaches while maintaining superior accuracy. Validation on the DHG-14/28 public benchmark confirms generalizability with 93.4 ± 1.6% accuracy. The framework shows robust performance under challenging conditions: low-light environments (92.4 ± 2.1%), fast motion (91.8 ± 1.9%), and complex backgrounds (93.2 ± 1.7%). Statistical significance was confirmed through comprehensive evaluation across diverse demographic groups. This research has implications for human-computer interaction applications, including smart home systems, augmented reality, and industrial automation.
This paper proposes the application of split computing paradigms for deep reinforcement learning through distributed computation between Connected Autonomous Vehicles (CAVs) and edge servers. While this approach has been explored in computer vision, it remains largely unexplored for reinforcement learning scenarios. We introduce a novel autoencoder trained directly through Deep Q-Network (DQN) rewards, wherein we optimize autoencoder layers using the DQN reward function while maintaining all other layers frozen. Our experimental results demonstrate that the proposed approach outperforms baseline methods by reducing data offloading requirements to the edge server by up to 98.7%. Additionally, this methodology not only decreases the data transmission burden but also achieves comparable rewards. In certain configurations, it even enhancing performance by up to 9.65%. The primary objective of this research is to reduce latency in deep reinforcement learning tasks for autonomous vehicles. In this regard, proposed approach achieves up to 66.5% improvement in latency reduction compared to baseline methods. These findings indicate that partial offloading through split computing offers significant benefits over both full offloading and complete on-device computation strategies for CAVs.
This paper compares two 3D sensing devices – a stereo camera and a Lidar. Both devices offer different modes, which perform differently, and have strengths and weaknesses. We have conducted a series of tests to assess the usability of these devices in our future research. The first round of tests focused on the interior performance and compared the various modes available. The tests were performed in good and bad conditions, showing best and worst-case scenarios. The second test round was performed outside with natural sunlight and more distant objects. The stereo camera was very susceptible to changes in light intensity and performed worse in the interior tests, but better in the exterior tests. The Lidar struggled with dark surfaces at closer ranges and during the exterior testing due to objects outside its range.
In recent years, we have witnessed the rapid development of artificial intelligence systems and their presence in various fields. These systems are very efficient and powerful, but often unclear and insufficiently transparent. Explainable artificial intelligence (XAI) methods try to solve this problem. XAI is still a developing area of research, but it already has considerable potential for improving the transparency and trustworthiness of AI models. Thanks to XAI, we can build more responsible and ethical AI systems that better serve people’s needs. The aim of this study is to focus on the role of the user. Part of the work is a comparison of several explainability methods such as LIME, SHAP, ANCHORS and PDP on a selected data set from the field of medicine. The comparison of individual explainability methods from various aspects was carried out using a user study.
This paper presents a system for autonomous indoor navigation of the DJI Mini 3 drone by combining an Android-based control application with ultra-wideband (UWB) technology for precise localization. The development process involved two main phases: (1) enabling drone control and sensor access through the DJI Mobile SDK and (2) integrating UWB-based positioning data for reliable indoor navigation. Three autonomous navigation methods were designed and evaluated, each leveraging UWB data to a different extent. The results demonstrate that incorporating UWB significantly improves positioning accuracy for our solution compared to relying solely on the drone’s built-in sensors. Overall, this work highlights the potential of UWB technology to enhance drone operations in environments where GPS is unavailable, paving the way for advanced applications in indoor logistics, monitoring, and other industrial or service-oriented scenarios.
Resilient data transmission in mobile networks is one of the biggest challenges that researchers and research groups are addressing. In mobile multi-hop network environments, different types of routing protocols are used for data transmission and routing, however, in Fifth-generation (5G) network environments, it is necessary to route and transmit data using protocols and methods that can respond to the specific characteristics and requirements of a given type of network. To this end, we propose a resilient data transmission method based on game theory and decentralized blockchain technology. The proposed method provides routing in the network using communities of nodes that are formed based on previous experience and Quality of Service (QoS) parameters. From the experimental simulations and the achieved QoS parameters, it can be analyzed that the proposed model can be an alternative method for routing and communication in networks without infrastructure, especially in situations when the network is attacked by different types of attacks and the impact of inappropriate mobile nodes on routing and communication, respectively.
This study explores the rapid advancement of Large Language Models (LLMs) and their growing support for languages beyond English. Despite this progress, the LLM ecosystem remains predominantly focused on English, particularly in key areas such as tokenizers and evaluation metrics. Traditional metrics like BLEU and ROUGE, designed for English, are less effective for languages with different linguistic structures, such as Slovak. This paper highlights the limitations of current metrics and the need for language-specific evaluation methods. The performance of three LLMs—Mistral 7B, Mistral 7B fine-tuned on Slovak, and Gemma 7B—was assessed using both human evaluators and automated evaluations by ChatGPT-4. Results show that fine-tuning on Slovak improves model performance, though errors in grammar and syntax persist. The study underscores the need for multilingual optimization and more inclusive AI evaluation tools to ensure effective language model performance across diverse linguistic contexts.
This paper addresses the integration of solar energy with electric vehicle charge points through a load management and distribution system. It proposes an algorithm that dynamically allocates available power, considering building consumption, electric vehicle charger needs, and real-time photovoltaic system output. The system ensures efficient power distribution and stability. Simulation results demonstrate the algorithm’s effectiveness in optimizing solar energy use, accommodating fluctuating power demands, and reducing costs, thus enhancing the sustainability and efficiency of EV charging infrastructure.
This article deals with the design of a generalized methodology for the identification of nonlinear dynamic systems using a combination of analytical and experimental identification methods. The proposed methodology is divided into modules that can be repeated if necessary. Verification of the methodology is presented on the modified helicopter educational model CE 150 from Humusoft company. A mathematical model of the system with both rotors approximated by a black-box nonlinear model is presented. The proposed experiments for obtaining experimental data are described in detail and are used in the estimation of model parameters. The gray-box model is validated in a control structure with a control law optimized for microcontrollers. Indirect validation of the gray-box model with the real helicopter educational model points to good approximation properties of the gray-box model and its potential for further use in the identification and control using artificial intelligence methods.
