Cross-Level Scene-Agent Graph Learning for Context-Aware Multi-Label Emotion Recognition
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171040 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von K. Suresh und 5 weiteren Autor:innen.
Abstract (Originalsprache)
In context-aware emotion recognition from images, it is necessary to consider the visual appearance of the target person along with the surrounding environment. In unconstrained settings, emotions may be simultaneously related to the target person, thus making the task multi-label. However, existing methods focus on either the representation of the target or the context without taking the interaction between the two into consideration. The current work proposes a scene-agent graph learning method for context-aware multi-label emotion recognition. An input image along with an annotation of the target person is passed through a pretrained ResNet-50 model to obtain 2048-dimensional visual representations of the scene and the target agent. Such representations are mapped to a 256-dimensional graph space where the scene and agent nodes are connected using an explicitly constructed relation structure and normalized graph propagation is performed. The proposed approach is validated on the EMOTIC dataset using the 26-category multi-label setting. The ablation study involves four different settings: target-only, scene-only, scene-agent reasoning, and full-fusion, using a consistent evaluation process. The Micro-F1 score of the target-only baseline is 0.260712, and its mAP is 0.256390, while the scene-agent reasoning approach improves Micro-F1 to 0.453180 and mAP to 0.275428. The best mAP is achieved by the full fusion setting with 0.278457, even though the Micro-F1 of 0.448038 is slightly lower than that of scene-agent reasoning. Category-level average precision, statistical paired analysis, error analysis at the sample level, and qualitative analysis provide additional insights into the performance of the proposed model.
Zitieren
APA
Suresh, K., Annabel, L. S. P., Venketbabu, T., Hariharasudhan, S., Sasikumar, M., & Ramar, A. (2026). Cross-Level Scene-Agent Graph Learning for Context-Aware Multi-Label Emotion Recognition. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171040
Vancouver
Suresh K, Annabel LSP, Venketbabu T, Hariharasudhan S, Sasikumar M, Ramar A. Cross-Level Scene-Agent Graph Learning for Context-Aware Multi-Label Emotion Recognition. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171040
BibTeX
@article{Suresh2026,
author = {Suresh, K. and Annabel, L. Sherly Pushpa and Venketbabu, T. and Hariharasudhan, S. and Sasikumar, M.S.S. and Ramar, A.},
title = {Cross-Level Scene-Agent Graph Learning for Context-Aware Multi-Label Emotion Recognition},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171040}
}
EUVIMED