Deep Learning-Based Touch-Event Recognition from Individual RGB Frames: A Comparative Study of CNN Architectures and Model Training Strategies
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171061 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Arash Ghazvini und 5 weiteren Autor:innen.
Abstract (Originalsprache)
Touch-event recognition from visual data remains a challenging computer vision task, particularly when distinguishing the body regions involved in physical interactions. Unlike video-based action recognition, frame-level touch-event recognition requires models to identify interaction classes from individual RGB images without relying on temporal motion information. However, the effectiveness of different convolutional neural network (CNN) architectures and training strategies for this specific recognition task requires further investigation. This study presents a comparative evaluation of six CNN architectures: ResNet50, ResNet18, EfficientNet-B0, MobileNetV3-Small, a ResNet18-based spatial-attention model, and a Simple CNN. The experiments use 815 unique RGB frames belonging to four classes: No Touch, Touch Head, Touch Hips, and Touch Shoulders. Five training strategies are investigated, including baseline training, targeted data augmentation, class-weighted cross-entropy loss, a combination of targeted augmentation and class-weighted loss, and weighted random sampling. To provide a consistent evaluation, five seeded participant-video-group partitions are employed, with model selection based on validation macro F1. Performance is assessed using accuracy, balanced accuracy, and macro F1. The results indicate that ResNet18 with class-weighted loss achieved the highest mean test macro F1 of 0.344, while the performance of training strategies varied across CNN architectures. These findings provide a comparative baseline for evaluating the influence of CNN architecture selection and training strategies on frame-level body-region touch-event recognition within the investigated dataset.
Zitieren
APA
Ghazvini, A., Hipiny, I., Amaran, M. A., Wahid, H. A., Abdullah, J., & Ujir, H. (2026). Deep Learning-Based Touch-Event Recognition from Individual RGB Frames: A Comparative Study of CNN Architectures and Model Training Strategies. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171061
Vancouver
Ghazvini A, Hipiny I, Amaran MA, Wahid HA, Abdullah J, Ujir H. Deep Learning-Based Touch-Event Recognition from Individual RGB Frames: A Comparative Study of CNN Architectures and Model Training Strategies. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171061
BibTeX
@article{Ghazvini2026,
author = {Ghazvini, Arash and Hipiny, Irwandi and Amaran, Mazdan Ali and Wahid, Hasnizam Abdul and Abdullah, Johari and Ujir, Hamimah},
title = {Deep Learning-Based Touch-Event Recognition from Individual RGB Frames: A Comparative Study of CNN Architectures and Model Training Strategies},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171061}
}
EUVIMED