Iryna Ahekian

Ms. Iryna Ahekian


Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 September 2019
Verified
Conference Paper
Closed access
OHOleh Hramm
Oleh Hramm
Nataliya BilousNataliya Bilous
Nataliya Bilous
IAIryna Ahekian
Iryna Ahekian
There are many methods for obtaining microscopic images, they can be obtained for different types of cells in different environments, which makes it impossible to create universal recognition algorithms. Therefore, for each specific task, it is necessary to implement a separate algorithm that will be effective for the existing set of images, and take into account the peculiarities of these images. The task of this work is the development of a flexible and customizable algorithm that can be configured to segment cells on different types of images. As a result, a solution was developed that has many customizable parameters to optimize the result for a specific data set. Also, this it is resistant to a lot of noise and artifacts that can occur on images, such as uneven background, small debris, loss of focus when shooting.

25 December 2025 in Applied System Innovation
Verified
Journal Article
Open access
PDF file
Nataliya BilousNataliya Bilous
Nataliya Bilous
Vladyslav MalkoVladyslav Malko
Vladyslav Malko
IAIryna Ahekian
Iryna Ahekian
Igor KorobiichukIgor Korobiichuk
Igor Korobiichuk
IVIVANICHEV VOLODYMYR
IVANICHEV VOLODYMYR
Reliable recognition of people on water from UAV imagery remains a challenging task due to strong glare, wave-induced distortions, partial submersion, and small visual scale of targets. This study proposes a hybrid method for human detection and position recognition in aquatic environments by integrating the YOLO12 object detector with optical-flow-based motion analysis, Kalman tracking, and BlazePose skeletal estimation. A combined training dataset was formed using four complementary sources, enabling the detector to generalize across heterogeneous maritime and flood-like scenes. YOLO12 demonstrated superior performance compared to earlier You Only Look Once (YOLO) generations, achieving the highest accuracy (mAP@0.5 = 0.95) and the lowest error rates on the test set. The hybrid configuration further improved recognition robustness by reducing false positives and partial detections in conditions of intense reflections and dynamic water motion. Real-time experiments on a Raspberry Pi 5 platform confirmed that the full system operates at 21 FPS, supporting onboard deployment for UAV-based search-and-rescue missions. The presented method improves localization reliability, enhances interpretation of human posture and motion, and facilitates prioritization of rescue actions. These findings highlight the practical applicability of YOLO12-based hybrid pipelines for real-time survivor detection in flood response and maritime safety workflows.

2 June 2020 in Bionics of Intelligence
Verified
Journal Article
Open access
PDF file
Nataliya BilousNataliya Bilous
Nataliya Bilous
ORO.V. Rassokha
O.V. Rassokha
IAIryna Ahekian
Iryna Ahekian
OHO.V. Hramm
O.V. Hramm
ІАІ.а Агекян
І.а Агекян
To determine the general state of human health using emotion recognition, a method based on machine learningwas chosen, the classifier is trained on the dataset “fer2013” and “PAB-F”. negative emotions and the use of a dataset totrain the neural network, with images of people divided into classes “painful” - “not painful”. For the “fer2013” dataset,it is necessary to determine the presence of pain after data processing. As a rule, pain is expressed in the intense andprolonged presence of emotions of anger and sadness. I suggest measuring the intensity of emotion by the probabilityfactor of determining the emotion, which returns the neural network. It has been experimentally determined that aspecialized data set better copes with the task, despite the fact that it has a sufficiently small number of images. “Fer2013”has a high percentage of false positives. This can be explained by the fact that the photos also showed obvious negativeemotions, which were regarded as pain. The number of learning epochs has a positive effect on the accuracy of theneural network, and the increase in learning speed has a negative effect. To check the accuracy of the neural network atthe entrance to the module to check the emotional state at the entrance were images found by the search query “Paingrimace” on the resource depositphotos.

1 June 2024 in IAES International Journal of Artificial Intelligence (IJ-AI)
Verified
Journal Article
Open access
PDF file
Nataliya BilousNataliya Bilous
Nataliya Bilous
OSOleh Svidin
Oleh Svidin
IAIryna Ahekian
Iryna Ahekian
Vladyslav MalkoVladyslav Malko
Vladyslav Malko
The aim of the work is to develop the method of identification and comparison of poses and exercises performed by a person that will have a low sensitivity to data errors. This method uses their formal descriptions in the form of conjunctions of logical statements and should work regardless of the shooting angle at which the video was taken and the proportions of the person on it. Each statement describes the position of the joints relative to each other along one of the axes. The joint coordinates are corrected by taking into account the length of the bones that connect them that eliminates the necessity to process outliers and it also improves the accuracy of joints positioning. Removal of errors out of the data using the method of averaging the graph along each axis at every step. In order to do this, consecutive points are grouped so that the difference between the maximum and the minimum does not exceed the error. The groups are then filtered to leave only those in which both are smaller or both are larger. The proposed method of identification requires just a modern smartphone and has no restrictions on how to take video of exercises.

18 May 2026 in Applied Sciences
Verified
Journal Article
Open access
PDF file
Nataliya BilousNataliya Bilous
Nataliya Bilous
Vladyslav MalkoVladyslav Malko
Vladyslav Malko
IAIryna Ahekian
Iryna Ahekian
MFMarcus Frohme
Marcus Frohme
The integration of artificial intelligence into video-based human behavior analysis enables contactless and continuous monitoring of both motor dynamics and facial reactions. This paper proposes a dual-stream multimodal framework for synchronized modeling of facial expression dynamics and skeletal motion during physical movement from monocular RGB video. The framework consists of two coordinated streams: the motor stream, based on 2D skeletal keypoints, and the facial stream, which extracts features associated with discomfort and affective responses. Person and face detection are performed using YOLO11, while specialized deep learning models handle pose estimation and facial expression recognition. Temporal dependencies and cross-modal relationships are modeled via a bidirectional LSTM, enabling unified temporal modeling of skeletal and facial dynamics. This novel approach allows investigation of how physical movement patterns relate to facial reactions during dynamic activities. By integrating heterogeneous facial and skeletal features in a synchronized temporal model, the framework enables consistent cross-modal analysis of dynamic human behavior. The framework was trained and validated using FER2013 and AffectNet for facial expression recognition, and UI-PRMD and FineRehab for skeletal motion modeling. It achieves 91.2% accuracy in facial expression classification, 94.8% mean Intersection over Union for human detection, and an F1 score of 0.89 for multimodal state assessment. Operating in real-time at 18–28 FPS on standard GPU hardware without requiring wearable sensors, the framework supports applications in behavioral monitoring and safety analysis.

1 July 2026 in Sensors
Verified
Journal Article
Open access
PDF file
Nataliya BilousNataliya Bilous
Nataliya Bilous
DODanylo Ostapchenko
Danylo Ostapchenko
IAIryna Ahekian
Iryna Ahekian
MFMarcus Frohme
Marcus Frohme
Remote tele-rehabilitation requires objective pain assessment, but existing approaches fail in two distinct ways. Self-report scales such as the Visual Analog Scale and the Numeric Pain Rating Scale are easy to falsify, opening a special case of the Oracle problem in blockchain-based insurance. Cloud-based computer vision handles falsification but transmits raw biometric video off the patient’s device, violating privacy requirements. A decentralized Edge AI-Oracle architecture is proposed that combines MediaPipe Face Mesh landmark extraction with a recurrent classifier mapping Action-Unit feature sequences to a learned pain score aligned with the Prkachin and Solomon Pain Intensity scale. The recurrent cell is selected empirically across short-context (T = 2) and long-context (T = 120 frames at 24 fps) regimes, with a two-layer Long Short-Term Memory (LSTM) network adopted for deployment. Inference and Elliptic Curve Digital Signature Algorithm (ECDSA) signing run inside an ARM TrustZone Trusted Execution Environment (TEE). Biometric logs are stored off-chain on the InterPlanetary File System (IPFS). Smart contracts anchor results on-chain and open a 24 h optimistic verification window for an off-chain Watchtower auditor. On SynPAIN the LSTM reaches F1 = 0.683 on T = 120 video (leave-one-stratum-out), with a directional but non-significant advantage over Gated Recurrent Unit (GRU) (Wilcoxon p = 0.167). Cross-dataset validation on BioVid Heat Pain Database Part A (87 subjects, 174 paired observations, leave-one-subject-out) yields F1 = 0.519 for LSTM and 0.499 for GRU (Wilcoxon p = 0.549). A processor-only TEE surrogate benchmark estimates 1.96 ms (FP32) and 0.45 ms (INT8) inference latency at T = 120 with a 0.34 MB footprint and 707 µs ECDSA signing latency, leaving the INT8 inference latency more than an order of magnitude below the 33 ms per-frame budget. The dual-layer storage reduces gas costs by a factor of 23.4 (160,261 vs. 3,744,872 gas), corresponding to an illustrative mainnet cost of approximately 0.53 USD per submission at 1 gwei, rising to roughly 16 USD at a busier 30 gwei, and falling to approximately 0.005 USD on Arbitrum One (April 2026 reference parameters), so that continuous monitoring is economically practical on Layer-2. An adaptive-adversary analysis of the Watchtower shows that gross score tampering is detected at every usable operating threshold, whereas a rational adversary who inflates by less than the dispute threshold, or who shapes the injected score to fall just inside it, evades detection. Because the false-positive rate reaches zero only for δ≳0.15, the protocol bounds rather than eliminates patient-side fraud and motivates a zero-knowledge proof-of-inference successor. The framework is architecturally and economically feasible as a cryptographically verifiable, privacy-preserving tele-rehabilitation substrate aligned with General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA) requirements through the Zero-Video Transmission principle, while remaining economically viable under post-Dencun mainnet and Layer-2 conditions. Recognition accuracy on real-world data and robustness to small-magnitude tampering remain limitations that the interchangeable recognition and audit components must improve before clinical deployment.