1. Efficient material selection for training occluded mmWave radar-based gesture recognisersNuwan Attygalle, Luis A. Leiva, Matjaž Kljun, Klen Čopič Pucihar, 2026, izvirni znanstveni članek Opis: Radar-based gesture recognition has emerged as a promising approach for unobtrusive interaction. Unlike camera-based systems, radar sensors can detect gestures through opaque materials, enabling seamless embedding into various everyday objects. However, it remains unclear how to train models efficiently for robust gesture recognition through diverse materials. To investigate this, we collected a dataset of 17,520 gesture recordings performed through 73 everyday materials. By comparing several material-sampling and data-augmentation strategies, we found that a small carefully selected representative subset of training materials was sufficient to match performance of a classifier trained on the full material dataset. Our results showed that the models trained on 14 quota-sampled materials achieved accuracies of 95.8% and 91.2%, comparable to training on all 73 materials (96.8% and 91.6%) and significantly better than training without material data (66.8% and 65.8%). Among the evaluated sampling approaches, Quota sampling also provided the best overall trade-off between performance and practicality. In contrast, classifiers trained on augmented data performed worse than those trained on actual material-specific data. Taken together, these findings indicate that, for the tested sensor, gesture set, and material collection, carefully selected real-material data offer a practical route to reducing material-specific data collection in radar-based gesture recognition while preserving generalisation. Code, models, and data are available in the public repository, with additional details provided in the supplementary materials: https://gitlab.com/hicuplab/seeing-through.
Ključne besede: deep learning, training optimization for neural networks, millimetre-wave radar, gesture recognition, material selection, sampling, synthetic noise
Objavljeno v RUP: 19.06.2026; Ogledov: 406; Prenosov: 12
 Celotno besedilo (15,56 MB)
Gradivo ima več datotek! Več... |
2. PN-SCA : a high generalization and fast profiled SCA based on prototypical networksYu Ou, Yongzhuang Wei, Changhai Ou, Enes Pašalić, 2026, izvirni znanstveni članek Opis: Recently, there has been a growing interest in scenarios that employ a few (imbalanced) power traces for deep learning based side-channel analysis (DL-SCA). Finding a lightweight DL-SCA architecture that is more generalizable and faster to learn is of great importance when handling such situations. In this work, an initial analysis is conducted on the capability of learning and the susceptibility to attacks of prevalent DL-SCA models, focusing on cases that culminate in an unsuccessful attack. Subsequently, a meta-learning technique, known as prototypical networks, is delineated for the construction of a lightweight deep learning framework. In contrast to conventional DL-SCA models, the proposed architecture, designated as PN-SCA, does not predict the probability of a power trace belonging to an intermediate value in classification. Instead, it facilitates the learning of an encoder capable of embedding power consumption data within a latent space, while also establishing templates, or prototypes, for diverse categories. Moreover, we developed a metric that is specifically intended for the selection of hyperparameters due to the unique training phase of PN-SCA. Finally, four distinct scenarios are established with few (including one imbalanced) power traces to evaluate the correctness of our architecture. The results clearly illustrate the advantages of our PN-SCA in terms of generalization, reduced training costs (with a decrease in profiling traces by over 50%), and significantly enhanced attack effect (with a reduction of power traces requirements by over 90%), thus demonstrating notable improvements over existing methodologies
Ključne besede: side-channel analysis, deep learning based SCA, prototypical networks, few-shot learning
Objavljeno v RUP: 20.05.2026; Ogledov: 477; Prenosov: 16
Celotno besedilo (3,78 MB)
Gradivo ima več datotek! Več... |
3. Image-based analysis of tourist destination perceptions : a deep learning and spatial–temporal study in SloveniaDejan Paliska, Aleksandra Brezovec, Gorazd Sedmak, 2026, izvirni znanstveni članek Opis: In the context of fierce competition among tourist destinations and increasing difficulty of differentiation, developing a strong destination image is particularly important. A comprehensive understanding of how tourists perceive destinations through user-generated images can help destination management organizations (DMOs) design more effective marketing strategies. This is especially relevant for destinations with spatially and temporally dispersed tourism resources and strong seasonal dynamics. This paper analyses inbound tourist photographs by combining deep learning techniques with spatial analysis to examine the spatial and temporal distribution of photo scenes and shifts in scene preferences among tourists. The study focuses on three distinct types of destinations in Slovenia—urban (Ljubljana), nature-based/alpine (Bled), and coastal (Piran, Izola, Koper)—providing insights into how image-based spatial scene analysis can inform destination marketing strategies. The results reveal significant spatial and temporal heterogeneity of scenes across micro destinations. Nature-based destinations exhibit lower topic entropy and fewer topic changes per user, whereas urban destinations show higher variability, with users changing topics on average five times per day. Seasonal effects are moderate: nature-based destinations display lower topic entropy in winter and higher in autumn and spring, coastal destinations show less pronounced seasonal variation, and urban destinations show almost none. These findings provide valuable insights into the spatial and temporal distribution of tourist interests and offer practical guidance for DMOs in strategic marketing planning.
Ključne besede: tourist destination image, user-generated content, deep learning, spatial-temporal analysis, destination marketing strategy
Objavljeno v RUP: 18.02.2026; Ogledov: 930; Prenosov: 14
Celotno besedilo (2,93 MB)
Gradivo ima več datotek! Več... |
4. Enhanced precision in axle configuration inference for bridge weigh-in-motion systems using computer vision and deep learningDomen Šoberl, Jan Kalin, Andrej Anžlin, Maja Kreslin, Klen Čopič Pucihar, Matjaž Kljun, Doron Hekič, Aleš Žnidarič, 2025, izvirni znanstveni članek Opis: Heavy goods vehicles (HGVs) have a significant impact on road and bridge infrastructure, with overloaded vehicles accelerating structural deterioration and increasing safety risks. Bridge weigh-in-motion (B-WIM) systems estimate gross vehicle weight (GVW) using strain measurements, but inaccuracies in axle configuration recognition can reduce reliability. This study presents a low-cost computer vision (CV) extension for existing B-WIM installations that verifies strain-inferred axle configurations using traffic camera images and flags GVW estimates as reliable or unreliable. Experiments on a data set of over 30,000 HGV records show that by combining convolutional neural networks with strain-based heuristics, GVW reliability can improve from 96.7% to 99.89%, effectively excluding nearly all erroneous measurements. The approach operates without interrupting ongoing B-WIM operations and can be applied retrospectively to historical data. Limitations include the inability to detect raised axles (RAs), which the method excludes as unreliable. This method provides a practical, high-precision enhancement for structural health monitoring of bridges.
Ključne besede: B-WIM, computer vision, deep learning
Objavljeno v RUP: 16.01.2026; Ogledov: 988; Prenosov: 8
Celotno besedilo (2,01 MB)
Gradivo ima več datotek! Več... |
5. Deep learning for brain MRI tissue and structure segmentation : a comprehensive reviewNedim Šišić, Peter Rogelj, 2025, pregledni znanstveni članek Opis: Brain MRI segmentation plays a crucial role in neuroimaging studies and clinical trials by enabling the precise localization and quantification of brain tissues and structures. The advent of deep learning has transformed the field, offering accurate and fast tools for MRI segmentation. Nevertheless, several challenges limit the widespread applicability of these methods in practice. In this systematic review, we provide a comprehensive analysis of developments in deep learning-based segmentation of brain MRI in adults, segmenting the brain into tissues, structures, and regions of interest. We explore the key model factors influencing segmentation performance, including architectural design, choice of input size and model dimensionality, and generalization strategies. Furthermore, we address validation practices, which are particularly important given the scarcity of manual annotations, and identify the limitations of current methodologies. We present an extensive compilation of existing segmentation works and highlight the emerging trends and key results. Finally, we discuss the challenges and potential future directions in the field.
Ključne besede: magnetic resonance imaging, brain, image segmentation, deep learning
Objavljeno v RUP: 10.10.2025; Ogledov: 1479; Prenosov: 24
Celotno besedilo (956,69 KB)
Gradivo ima več datotek! Več... |
6. |
7. |
8. A lightweight deep learning model for profiled SCA based on random convolution kernelsYu Ou, Yongzhuang Wei, René Rodríguez, Fengrong Zhang, 2025, izvirni znanstveni članek Opis: In deep learning-based side-channel analysis (DL-SCA), there may be a proliferation of model parameters as the number of trace power points increases, especially in the case of raw power traces. Determining how to design a lightweight deep learning model that can handle a trace with more power points and has fewer parameters and lower time costs for profiled SCAs appears to be a challenge. In this article, a DL-SCA model is proposed by introducing a non-trained DL technique called random convolutional kernels, which allows us to extract the features of leakage like using a transformer model. The model is then processed by a classifier with an attention mechanism, which finally outputs the probability vector for the candidate keys. Moreover, we analyze the performance and complexity of the random kernels and discuss how they work in theory. On several public AES datasets, the experimental results show that the number of required profiling traces and trainable parameters reduce, respectively, by over 70% and 94% compared with state-of-the-art works, while ensuring that the number of power traces required to recover the real key is acceptable. Importantly, differing from previous SCA models, our architecture eliminates the dependency between the feature length of power traces and the number of trainable parameters, which allows for the architecture to be applied to the case of raw power traces.
Ključne besede: side-channel analysis, deep learning, convolution neural networks, random convolution kernel
Objavljeno v RUP: 26.09.2025; Ogledov: 2149; Prenosov: 15
Celotno besedilo (1,75 MB)
Gradivo ima več datotek! Več... |
9. |
10. |
