StressIRNet: A Novel Lightweight CNN Architecture for Stress Classification Leveraging Smartphone Thermal Imaging Modality
Artykuł w czasopiśmie
MNiSW
100
Lista 2024
| Status: | |
| Autorzy: | Baran Katarzyna |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2025 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 13 |
| Strony: | 194475 - 194488 |
| Impact Factor: | 3,6 |
| Web of Science® Times Cited: | 0 |
| Scopus® Cytowania: | 0 |
| Bazy: | Web of Science | Scopus | |
| Efekt badań statutowych | NIE |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | TAK |
| Licencja: | |
| Sposób udostępnienia: | Otwarte czasopismo |
| Wersja tekstu: | Ostateczna wersja opublikowana |
| Czas opublikowania: | W momencie opublikowania |
| Data opublikowania w OA: | 13 listopada 2025 |
| Abstrakty: | angielski |
| The pervasive and growing incidence of stress-related disorders underscores a critical need for accessible, non-invasive monitoring technologies. Conventional physiological modalities such as electroencephalography (EEG), electrocardiography (ECG), and galvanic skin response (GSR), while providing valuable data, are often constrained by their cost, obtrusiveness, and lack of suitability for continuous, daily use. This work posits that the synergistic combination of smartphone-based thermal imaging and purpose-built lightweight deep learning architectures presents a viable and compelling alternative. This work introduces StressIRNet, a novel lightweight convolutional neural network (CNN) specifically engineered for the task of human stress classification from facial thermal imagery captured via mobile thermal cameras. The architectural design of StressIRNet is a hybrid that strategically integrates three key components: Ghost Modules for efficient, redundancy-reduced feature generation; Channel Shuffle operations to facilitate robust cross-group information flow amongst features; and a Squeeze-and-Excitation (SE) mechanism that performs dynamic, channel-wise feature recalibration. This design philosophy prioritizes a Pareto-optimal trade-off between discriminative performance and computational footprint. Extensive experimental evaluation on a proprietary dataset (StressIR), comprising 120 subjects, demonstrates that StressIRNet achieves a leading accuracy of 77% in a four-class stress classification problem (No Stress, Low, Medium, High). Crucially, in a comparative analysis against state-of-the-art lightweight CNNs including MobileNetV2, MobileNetV3-Small, EfficientNet-B0, and ShuffleNetV2, StressIRNet not only achieves superior accuracy but does so with the lowest computational complexity, requiring only 1.05 million parameters and 0.04 GFLOPs. External validation on two public thermal imaging benchmarks confirmed the model’s generalization capability, yielding a robust accuracy of 72%. These results collectively affirm the high potential of the proposed solution as a foundational technology for practical, real-time, and energy-efficient stress monitoring systems in mHealth and affective computing applications. |
