%T  International Conference on User-System Interaction - ICUSI 2025
%C  Iasi
%D  18-19 September 2025
%G  ISSN 3061-3205
%E  Alex Moruz
%E  Cesar Collazos
%P  181
%W  https://rochi.utcluj.ro/icusi2025

%T Accessibility of local government websites
%A Costin Pribeanu
%P 1-4
%X Although European regulations require web accessibility of public sector bodies by June 2025, the accessibility of municipal websites remains low. This paper reports an accessibility evaluation that targeted the first 100 Romanian websites. Three accessibility checking tools were used in this study: Total Validator, Wave, and AChecker. Although the comparison with previous data from 2019 shows a slight improvement, web accessibility remains low, with few websites meeting the requirements of WCAG 2.0. Without clear regulations at the national level, meeting the demands of European regulation is problematic.
%K Web accessibility, WCAG2, accessibility checking tools, municipal websites.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p1-4.pdf

%T Interactive Tool for Sign Language Alphabet Acquisition
%A Rebeca Costache
%A Adrian Iftene
%P 5-12
%X Sign language education remains inadequately supported by existing interactive learning platforms, often lacking inclusive and engaging experiences. This paper presents a human-centered application aimed at addressing this problem by employing the Kinect, Microsoft’s gesture-based sensor, to facilitate immersive sign language instruction. Furthermore, integrating the Kinect with Machine Learning (ML) algorithms can boost the pedagogical progression in sign language acquisition, as learning is most efficient when theory is combined with active practice. In the development process of this application, various models were trained on three different datasets in order to obtain the best performance in learning and predicting the American Sign Language (ASL) alphabet. By integrating Human-Computer Interaction (HCI), this application represents the first step in supporting not only members of the Deaf and Hard of Hearing (DHH) communities, but also individuals from broader linguistic, educational, and professional backgrounds who seek to learn sign language for communication and education, or to advocate for Deaf rights and accessibility.
%K Gesture-based interaction; Kinect; neural networks; educational technology; American Sign Language; humancomputer interaction.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p5-12.pdf

%T LetterHeroes: Designing Support Systems for Improving Literacy Levels in Romanian Children with Dyslexia
%A Viktoria Boda
%A Adriana-Mihaela Guran
%A Sorina-Georgeta Groza
%A Grigoreta-Sofia Cojocar
%P 13-20
%X Dyslexia, also called specific reading disability, is a neurodevelopmental disorder that characterizes moderate and imprecise word identification and recognition, impairments in decoding, word reading accuracy and fluency, and spelling. In Romania, awareness and recognition of dyslexia as a distinct learning difficulty is relatively recent and effective digital support tools in Romanian are lacking. Dyslexic children need adapted support in letter recognition, syllable identification and pronunciation to develop their reading ability. We describe a user-centered design process in developing an application that supports early literacy development (letter recognition/identification, phoneme identification, syllables, spelling and reading) for the Romanian language. We describe the scientific foundation for our approach, the design process and the results of qualitative evaluation with field experts.
%K Dyslexia; assistive technology; User Centered Design
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p13-20.pdf

%T Interactive Virtual Reality Scenarios for Treating Phobias: Enhancing User-System Interaction in Mental Health
%A George-Gabriel Constantinescu
%A Adrian Iftene
%P 21-26
%X In the context of immersive technologies in healthcare, virtual reality (VR) has been identified as a promising tool for treating phobias through controlled and interactive exposure scenarios. This paper presents various VR-based approaches for addressing specific phobias by simulating real-life situations in safe and user-friendly environments. Using interactive design techniques and natural user-system interaction models, such as gaze-based triggers, hand-based interactions, and scenario adaptation, our study enhances the therapeutic experience. It aims to reduce patients' anxiety responses over time. The system enables repeated gradual exposure and dynamic difficulty adjustment, making therapy more accessible, customizable, and engaging. Our research emphasizes the role of VR as a complementary tool in mental health care, capable of preparing users to better manage different phobias in everyday life. The study further highlights the potential of user-centered interactive technologies to support long-term behavioral improvements in clinical and home-based therapeutic contexts.
%K Virtual Reality (VR); Phobia Treatment; Exposure Therapy; User Experience (UX); Human-Computer Interaction (HCI); Interactive Scenarios; Mental Health; User-System Interaction
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p21-26.pdf

%T Usability and Accessibility of Romanian Online News Websites
%A Marian Pădure
%A Costin Pribeanu
%P 27-32
%X The usability and accessibility of online news websites seem to have been neglected in the last decade. Few studies exist that check the accessibility for blind and visually impaired people, and even fewer the usability of online news websites. The objective of this paper is twofold. First, it is to explore the usability of online news websites using a feature-based approach. The second is to analyze accessibility against WCAG 2.0 using a tool-based approach. Three accessibility checking tools have been used: Wave, AChecker, and TAW. The sample consisted of 18 Romanian online news websites. The evaluation results show many usability issues and low web accessibility.
%K Online news websites, usability, accessibility, WCAG2, and checking tools.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p27-32.pdf

%T Enhancing Developer Comprehension of Error Notifications through Visual Aid
%A Stefano Casafranca
%P 33-41
%X This paper investigates how software developers understand compiler error messages and proposes visual annotations to support their comprehension. The research addresses the limitations of current Integrated Development Environments (IDEs) in facilitating developers’ self-explanation processes. A user study with 28 software engineering students demonstrates that the proposed annotations improve developers’ ability to interpret error notifications, leading to more accurate mental models.
%K Compiler errors; Visualization; Self-explanation; User-centered design; IDE feedback
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p33-41.pdf

%T SimulEase: A VR-Assisted Approach to Overcoming Social Anxiety
%A Delia-Elena Ungureanu
%A Adrian Iftene
%P 43-50
%X The aim of this project and thesis is to develop and analyze the results of SimulEase. This virtual reality (VR) program replicates real-life situations that cause anxiety in a controlled, repeatable, and safe setting. Through immersive exposure and AI stress detection (from audio and text user input), the software seeks to assist users in better managing the symptoms of social anxiety and stress associated with public speaking. The participants can become tolerant and learn coping behaviors in a step-by-step process without having to endure the consequences of their behavior in real life by mimicking anxiety-evoking events in a well-controlled, risk-free VR environment. By putting forth an interactive solution that combines immersive technology with real-time emotion recognition for individualized exposure therapy and self-guided support, our study seeks to advance the expanding field of digital mental health solutions. Research shows that VRET (VR Exposure Therapy) is a promising supplementary treatment for SAD (Social Anxiety Disorder), especially when combined with CBT (Cognitive Behavioral Therapy).
%K Virtual reality, Exposure therapy, Social anxiety disorder
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p43-50.pdf

%T BioLabVR: A Virtual Reality Application for Interactive Learning with Student Engagement Analytics
%A Gesica Goron
%A Adrian Sabou
%P 51-58
%X This paper presents BioLabVR, an immersive educational system using Virtual Reality on Meta Quest 2 to improve the way students interact with biological concepts. Integrating hand-tracking, the application enables learners to navigate detailed 3D anatomical models, examine microscopic structures, and engage in quizzes aligned with curriculum objectives. Developed in Unity and enhanced with Cognitive3D analytics, the platform integrates intuitive interaction with behavioral data collection, offering insights into learner engagement and performance. By shifting from passive content delivery to hands-on exploration, the system enhances cognitive immersion, sustained attention, and a stronger connection between students and the concepts they study, encouraging deeper understanding in a modern educational context.
%K Virtual Reality; immersive learning; anatomy education; learning analytics; educational application
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p51-58.pdf

%T Flower Identifier – Augmented Reality Application for Flower Recognition
%A Constantin-Sebastian Bolohan
%A Adrian Iftene
%P 59-65
%X This paper presents a mobile application designed to
identify flowers in real time using a combination of deep
learning and augmented reality. The system integrates a
MobileNetV2 model trained via transfer learning on the
Oxford 102 Flowers dataset, enhanced through aggressive
image augmentation techniques. The application is built
using Unity and AR Foundation for cross-platform support,
and uses Firebase for authentication and real-time data
management. A chat bot powered by the Gemini API offers
botanical explanations in natural language. The app
supports live flower recognition through the camera, AR
display of relevant information, a personal flower
collection, and an interactive assistant. Usability testing
revealed a high level of user satisfaction, with an average
rating of 4.6/5. The results demonstrate the feasibility of
combining AI and AR to create an engaging, educational,
and mobile-friendly experience.
%K Augmented Reality; Mobile Applications; Flower Recognition; Deep Learning; MobileNetV2; Unity; Firebase
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p59-65.pdf

%T Interactive Development of a Virtual Reality System for Training and Evaluation in Electric Vehicle Technology
%A Silviu Mihaita Pescaru
%A Rafaella Vlase
%A Dorian Gorgan
%P 67-74
%X The global transition to electric mobility necessitates a fundamental recalibration of professional skills in the automotive sector. The complexity of high-voltage systems, battery packs, and software diagnostics in electric vehicles (EVs) creates a significant skills gap for technicians. Traditional training methods, which involve costly physical prototypes and inherent safety risks, are becoming increasingly inefficient. This paper proposes the development of a software system, called VRQUIZ, based on virtual reality (VR) designed to facilitate a complete training and evaluation cycle, from the creation of interactive training modules to the practical simulation of diagnostics and repairs, and finally, to performance evaluation. The system consists of an intuitive desktop application for professors, an immersive VR application for students, and a centralized server. The paper analyzes experimental solutions of user interaction techniques in a virtual reality system for training and assessing knowledge in electric vehicle technologies.
%K Electric Vehicles, Virtual Reality, User Interaction Techniques, Training Lessons.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p67-74.pdf

%T ARtist - Educational Experience in Art through Augmented and Virtual Reality
%A Maria Ciorâțanu
%A Adrian Iftene
%P 75-82
%X ARtist is a mobile application that integrates Augmented Reality (AR) and Virtual Reality (VR) to support drawing practice and art education. The AR component overlays drawing guides on physical paper, while the VR component presents a virtual museum containing forty famous paintings and contextual information. The system, implemented in Unity for Android, comprises eight modules, including step-by-step drawing tutorials, daily reference images, an AI-based art tutor using OpenAI API, a personal gallery, the possibility to place art on your real walls using AR, and a knowledge quiz. An on-device CNN estimates the difficulty of any reference image and allows users to adapt accordingly. A formative usability study with sixteen participants (aged 13–55) awarded the application the maximum overall score, while recommending minor refinements that were subsequently incorporated. The paper details the system architecture, the integration of AR and VR workflows into a coherent learning path, and the main findings of the user study, illustrating the potential of computer science to widen access to cultural education and the fact that innovation should not replace art, but enrich it and bring it closer to people.
%K Augmented Reality; Virtual Reality; Artificial Intelligence
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p75-82.pdf

%T Interactive Visualisation of AI Clustering Algorithms with VR Integration
%A Anda-Teodora Cosma
%A Dorian Gorgan
%P 83-90
%X Understanding certain abstract concepts, as well as the operational mechanisms of specific algorithms, is a difficult task. Dynamic visualization of these algorithms in action is a real challenge, especially when trying to distinguish the differences and similarities between algorithms in the same category, such as clustering algorithms. This research aims to bridge this gap by developing an interactive tool that employs Virtual Reality (VR) technology and the Meta Quest 2 headset to facilitate a more enjoyable and engaging learning experience. It explores solutions that integrate a classification algorithm with key clustering algorithms in Artificial Intelligence (AI) to visualize and explore hidden patterns in multidimensional data in a 3D virtual space. Experiments highlight the issues and benefits brought by a certain level of usability of user interaction techniques.
%K Virtual Reality, Classification Algorithm, Clustering Algorithm, Visual Analysis, Multidimensional data.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p83-90.pdf

%T CarDrivingTutor - A Multi-Modal Framework for Answering Car Driving Exam Questions
%A Stefan-Paul Buracu
%A Laurentiu-Marian Neagu
%A Mihai Dascalu
%P 91-98
%X Successful preparation for driving license examinations requires accurately interpreting both textual information and visual cues to represent realistic driving situations. Current systems for answering driving license exam questions in Romania predominantly rely on textual understanding. This paper introduces CarDrivingTutor - a multimodal framework for answering multiple-choice driving license questions in Romania, integrating a Retrieval-Augmented Generation (RAG) strategy using Chroma DB with semantic search, and YOLOv8 for road sign detection. To study the knowledge of LLMs in this area, two models, Gemma3:27B and LLaVA:34B, are analyzed under 3 scenarios ranging from a simple version where the model receives only the question and the answer options, a version in which each question is accompanied by relevant legal context retrieved from the Road Safety Legislation through semantic search, and a version where the textual context from the previous point is also accompanied by official definitions extracted by automatic detection of traffic signs from images. Results show that added textual context significantly improves accuracy, especially for Gemma3. Its accuracy increased from 52% to 64% for questions without images. In contrast, LLaVA showed smaller gains, from 46% to 51%. The use of visual definitions improved Gemma3’s accuracy slightly (from 46% to 51%), while LLaVA’s performance remained unchanged.
%K Multi-modal Framework; Driving License Exam; Large Language Models; Retrieval-Augmented Generation; Semantic Search; Traffic Sign Recognition; YOLO
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p91-98.pdf

%T Comparison of Python code generated by ChatGPT and AlphaCodium
%A Rareș Mihai Dinu
%A Marian Cristian Mihaescu
%A Traian Eugen Rebedea
%P 99-106
%X This paper presents a comparative analysis of the performance of ChatGPT and AlphaCodium in generating Python code for algorithmic problem-solving, using tasks sourced from the competitive programming platform Codeforces. The study evaluates how effectively each model interprets problem statements, generates correct and efficient solutions, and handles edge cases and complex scenarios under the platform's strict time and memory constraints. Both models were assessed by executing their generated solutions, which were then evaluated by Codeforces' judging system. Unlike previous studies that often rely on closed or controlled environments, this research leverages an open, reproducible, and objective testing framework provided by Codeforces to deliver quantifiable and verifiable results. The findings contribute to a deeper understanding of the practical capabilities and limitations of AI-based code generation models in competitive programming contexts, offering valuable insights for both academic research and industry applications.
%K AI Code Generation, ChatGPT, AlphaCodium, Competitive Programming, Python, Codeforces, Algorithmic Problem Solving, Natural Language Processing, Program Synthesis, Model Evaluation
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p99-106.pdf

%T Analysis of persona assigned LLMs
%A Ioana-Călina Pascu
%A Ștefan Trăușan-Matu
%P 107-114
%X As generative language models become more capable of imitating authorial style, traditional ideas about originality and literary expressiveness must be reconsidered. This paper aims to evaluate, through a computational approach, the extent to which a large language model (LLM) can reproduce the stylistic patterns of two well-known authors: James Joyce and E.M. Forster. The comparative analysis relies on the extraction of features grouped into three categories: stylometric (statistics on structure and vocabulary), idiolectal (unusual words, neologisms, sentence types), and rhythmic-prosodic (sound-based rhetorical devices). These features are examined in both original and machine-generated texts using quantitative measures such as Euclidean distance, cosine similarity, Jensen–Shannon divergence, and the Jaccard index. The results reveal subtle but significant differences between what can be algorithmically reproduced and what remains unique to human literary expression.
%K stylometry, idiolect, rhetorical devices, generative AI
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p107-114.pdf

%T Explaining Explanations using Large Language Models
%A Vlad Lungu
%A Ștefan Trăușan-Matu
%P 115-122
%X With deep neural networks being steadily adopted in various sectors such as healthcare, finance or tech, the need for interpretability of their results becomes paramount. In this context, not only is it difficult to produce reliable explanations of such complex models, but even more importantly, non-expert users - the main beneficiaries of real-world applications of artificial intelligence (AI) systems - may not have the capacity to comprehend and make use of these explanations. The contributions related to this paper are twofold. Firstly, we extend the work of Zytek et al. [19] by generating explanations for 3 different explainability techniques on the IMDB dataset [9], using different prompts. Secondly, we analyze the explanations produced to draw insights about the behavior and potential biases of the underlying sentiment classification model.
%K explainable artificial intelligence; natural language processing; deep learning; transformers; prompt engineering
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p115-122.pdf

%T Digital Technology Integration in the Educational Practice - An Exploratory Study
%A Gabriel Gorghiu
%A Elena-Ancuța Santi
%A Costin Pribeanu
%P 123-126
%X In the context of the accelerated digitization of the education system, the teaching process is facing a major transformation, marked by multiple opportunities as well as challenges that require continuous teacher training, adaptability, investment in infrastructure, and coherent educational policies. The contemporary teaching approach requires teachers to reconfigure traditional pedagogical paradigms to create relevant learning contexts focused on the specific educational needs of digitally native students. The main objective of this research is to analyze how the ease of integrating digital technology into the classroom and its perceived usefulness influence its effective use in teaching. The results show that both perceived ease of use and usefulness are important factors that determine the effective use of technology in the teaching process.
%K Digital technologies in education, technology acceptance, Continuous Professional Development Program, and online teaching.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p123-126.pdf

%T Advanced Activity Analysis for Tesys e-Learning platform
%A Gabriela-Cristiana Costache
%A Paul-Stefan Popescu
%A Marian Cristian Mihaescu
%A Costel Marian Ionașcu
%P 127-134
%X This paper presents the development and integration of an advanced activity analysis module into the Tesys e-learning platform. Motivated by accreditation requirements and a need for deeper insights into user behavior, the project aimed to provide comprehensive statistics and visualizations of students and faculty engagement. The module was implemented with minimal disruption to existing infrastructure, offering intuitive dashboards, role-based filtering, and data export capabilities. Using anonymized real-world data, the system was evaluated through visual analytics and statistical modeling. Results show significant variance in user engagement and a positive correlation between platform activity and assignment submission. The tool enables evidence-based decision-making for administrators and enhances the overall pedagogical value of the platform.
%K Activity Analysys; Human Computer Interaction; E-learning platform; Tesys
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p127-134.pdf

%T Developing a comprehensive Augmented Reality-based Software Platform for Educational Embedded Systems
%A Vlad Bartolomei
%A Victor Ioan Bâcu
%P 135-140
%X As the number of embedded system-devices has an exponential growth in the recent years, the works in-behind are still hard to teach and illustrate by teachers. 2010s have brought augmented reality in the game, a new technology whose main qualities are visualisation enhancement in real environment and interactability with the objects. We propose combining of two of the most used technologies in today’s world by using for illustration purposes the best documented board from Arduino by visually expanding the datasheet: the Uno R3. We developed a software platform based mainly on visualisation with apps for both teacher and student so that a sense a collaboration between these two entities would be still possible.
%K Augmented Reality, Model Target, Image Target, Vuforia, Engineering Education, Arduino Uno
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p135-140.pdf

%T Cognitive Load Detection Using Physiological Data from Commercial Smartwatches
%A Raul-Alexandru Gorgan
%A Teodor-Traian Ștefănuț
%P 141-148
%X The detection and management of cognitive load has become a priority in today’s technology-based world, where mental effort impacts performance and health quality. Traditional assessing methods rely on expensive and intrusive laboratory equipment, limiting the real-world applicability. This paper presents a non-invasive system for cognitive load detection using only physiological signals from commercial smartwatches. The proposed architecture includes a smartwatch application for real-time sensor data acquisition, a server for processing and storage and a web interface for visualization and control. An experimental methodology with alternating cognitive and relaxation phrases, inspired by the Pomodoro technique [1], was used to induce physiological responses. Key metrics such as heart rate (HR), heart rate variability (HRV), skin temperature, photoplethysmography (PPG) and task performance were analyzed. Results show significant HRV and HR variations between cognitive and non-cognitive phases, confirming the feasibility of wearable-based workload assessment. This work demonstrates that lowcost, accessible smartwatches can support cognitive monitoring, with using in educational, professional or health-related contexts.
%K Cognitive Load Detection; HRV; Physiological Data; Commercial Wearables; PPG.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p141-148.pdf

%T Automatic Emotion Recognition from Videos. Does the video file format matter?
%A Ionut-Daniel Morjan
%A Grigoreta-Sofia Cojocar
%P 149-156
%X Emotion recognition from video is an important capability for applications like driver monitoring or telemedicine. Although the latest state-of-the-art approaches achieve good accuracy, there is a gap in presenting these results. How do they perform if the video file quality is degraded by variation in resolution or compression rate, or codec? In this work, we compare two state-of-the-art approaches, Poster++ and DAN, by measuring their robustness when the video file quality is degraded. We used a high-quality MP4 sequence as our reference and we created versions that vary in resolution, compression bitrate, and codec. The results obtained showed that under moderate conditions, both approaches maintained consistency close to their high-quality baseline, with most frames yielding identical emotion labels, but when the video quality dropped further, differences emerged. At low resolution with heavy compression, Poster++’s accuracy fell into the mid-eighties for modern codecs, while DAN stayed above the low-nineties for those same inputs. An additional interesting result observed is that both models remained highly confident even when many frames were misclassified, indicating that confidence alone cannot detect poor input quality. Overall, our study demonstrates that while recent emotion recognition approaches can tolerate mild to moderate video degradation, severe downsampling or heavy compression can significantly compromise their performance.
%K automatic emotion recognition, video quality
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p149-156.pdf

%T Star ratings as a predictor for sentiment polarity in online reviews - a case study
%A Adriana-Mihaela Guran
%A Grigoreta-Sofia Cojocar
%A Dan Cojocar
%P 157-160
%X Nowadays, the web is abundant in information related to product and services. People are using online reviews to support their decisions. Sometimes, consumers focus their attention on the star rating of a product, which is a quantitative assessment of the product, while the text of the review provides qualitative information, but more time consuming to read and process. Still, there are few studies verifying the consistency between the star rating of a product or service and the sentiment of the review, and no study for Romanian reviews. In this paper we discuss an approach on verifying the consistency between star ratings and sentiment polarity for a set of Romanian reviews gathered from an e-commerce site. We measure the consistency between the two series and the results show a good consistency for our data.
%K sentiment analysis, star ratings, customer reviews
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p157-160.pdf

%T An Analysis of Rubik's Cube Solving Algorithms
%A Anca Ionela Nicu
%A Paul-Stefan Popescu
%P 161-168
%X Nowadays, there are numerous methods to solve the Rubik’s Cube of various dimensions, most of which aim to find the optimal solution. Our application is an addition to them, focusing on discovering a solution to the pocket Rubik’s Cube by applying six different algorithms and comparing their efficiency. Each of the six methods has been run four times on a set of 45 test data (five cube configurations for each optimal solution length from one to nine, where the optimal solution represents the shortest sequence of moves that can solve a cube), and for each test data, we computed the success rate in discovering a solution, the average memory used, the average execution time, and the average solution length (the average number of moves a solution has). Based on these results, we grouped the findings by the optimal solution length and compared the algorithms on the four criteria mentioned earlier. Comparing the algorithms' results, we find that the Layer-by-layer method is the best in terms of memory, execution time and success rate, but the worst in terms of solution length, A* and bidirectional BFS (Breadth-first search) are good overall for each criterion considered, DFS (Depth-first search) has the lowest success rate overall, and MCTS (Monte Carlo Tree Search) and Qlearning do not perform well, especially on complex cube configurations (cubes that can be solved optimally on six to nine moves). This research contributes to the field of Rubik’s cube solvers by developing six different ways from scratch to solve the 2x2 cube and by demonstrating their efficiency on different cube configurations, in comparison to each other.
%K Rubik’s cube, algorithms comparison, bidirectional BFS, DFS, MCTS, Q-learning, A*, Layer-by-layer
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p161-168.pdf

%T SkyWeather -cloud classification on mobile devices
%A Stefan Alexandru Glava
%A Paul-Stefan Popescu
%A Marian Cristian Mihaescu
%P 169-174
%X SkyWeather turns an ordinary Android phone into a pocket “sky-scanner” that recognises five cloud types (clear, veil, patterned, thick-white, thick-dark) completely offline, then offers short-term weather insight in under three seconds. The application couples a 6.9 MB MobileNetV2-based classifier, quantised with TensorFlow Lite, with a multilingual Jetpack- Compose interface that runs on mid-range devices. A hybrid dataset, SWIMCAT, which comprises more than 290 field photos captured across seventeen geo-climates, drives training. Targeted augmentation and class-weighted loss push validation accuracy to 94.7%, while keeping inference latency at 1.7 seconds end-to-end. Experiments confirm robust performance across varying daylight, altitude, and sensor noise conditions, with < 10 % confusion between visually similar classes. By eliminating server calls, SkyWeather ensures privacy, autonomy, and resilience for hikers, farmers, and first responders operating beyond network coverage. The project contributes (i) an open, mobile-ready cloud image corpus, (ii) an edge-optimized CNN pipeline, and (iii) a clean-architecture Android reference implementation, demonstrating that practical, citizen-oriented meteorology can live entirely on-device.
%K Cloud classification; machine learning; mobile app; offline inference; TensorFlow Lite; Kotlin; Android; humancomputer interaction.
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p169-174.pdf

%T Plant Recogniser: Offline Plant Identification on Android
%A Constantin Dan Foamete
%A Paul-Stefan Popescu
%A Marian Cristian Mihaescu
%P 175-180
%X This paper presents Plant Recogniser, an Android application that performs real-time plant identification entirely offline, thereby eliminating the network latency and privacy concerns typically associated with cloud-based solutions. The system combines a curated, class-balanced dataset (143 species; 18,753 test images) with a two-phase fine-tuning pipeline that adapts MobileNetV2 to the botanical domain. Post-training int8 quantisation produces a 14.8 MB TensorFlow Lite model that runs with a median latency of 1.3 s on mid-range Snapdragon 778G devices, without loss of accuracy. On the held‑out test split, the model attains 80.37 % top‑1 accuracy, 0.73 macro‑precision, and 0.71 macro‑F1, matching heavier EfficientNet‑Lite0 baselines while using 60 % less storage and energy. A user‑centred Jetpack Compose interface surfaces top‑3 predictions with coloured confidence bars, integrates optional Firebase history logging, and remains fully functional offline. Field tests with horticulture students confirmed a 91% perceived correctness rate and task completion within < 4 taps. These results demonstrate that lightweight, edge-ready CNNs, when paired with thoughtful interaction design, can deliver classroom- and field-grade biodiversity assistance without sacrificing speed or battery life.
%K plant recognition; mobile AI; Android; TensorFlow Lite; MobileNetV2; biodiversity; image classification; Jetpack Compose
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p175-180.pdf

%T Lightweight Day and Night Pedestrian Detection for Sensitive Driver Assistance
%A Alexandru-Rafael Hodor
%A Paul-Stefan Popescu
%A Marian Cristian Mihaescu
%P 181-187
%X Pedestrians account for a disproportionate share of urban traffic casualties, most of which occur at speeds below 50 km/h, where timely driver warning could avert impact. We present a low-cost, camera-only Advanced Driver- Assistance System that detects, tracks, and range-estimates pedestrians in real time on commodity hardware. The core detector is YOLO11n1 (≈ 4 M parameters) fine-tuned on a 50-50 day-night subset of the EuroCity Persons dataset2 (4 650 day and 4,222-night training images), eliminating the daylight bias typical of existing models. The system sustains a latency of 33 ms per 640 × 480 frame (<1 GB RAM). A ByteTrack module maintains identities across frames, and a pin-hole projection coupled with a time-to-collision filter triggers visual alerts when braking distance is insufficient. Unlike prior work that assumes discrete GPUs, multispectral sensors or daylight scenes, our pipeline delivers illumination-robust detection, constant-frame-rate inference and modular open-source implementation on hardware already present in budget vehicles. The results demonstrate that adequate pedestrian warning can be achieved without specialised sensors or expensive accelerometers, paving the way for wider deployment in cost-sensitive markets and retrofit scenarios.
%K Machine learning; object detection; YOLO; ECP; pedestrians
%W https://rochi.utcluj.ro/icusi/proceedings/articles/2/ICUSI_2025_p181-187.pdf