Design and Performance Analysis of an Efficient Deep Learning Architecture for Facial Emotion Recognition
International Journal of Drug Delivery Technology · 29.05.2026 · 16 (40s)
DOI: 10.25258/ijddt.16.40s.87 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 29.05.2026. Verfasst von Abhishek Gulia und 1 weiteren Autor:innen.
Abstract (Originalsprache)
The rapid advancement of Facial Expression Recognition (FER) has positioned deep learning as the dominant paradigm for interpreting human affective states in real-world applications such as Intelligent Transportation Systems (ITS) and Advanced Driver Assistance Systems (ADAS). However, a critical challenge persists: existing high-capacity models achieve impressive accuracy at the cost of computational redundancy, rendering them incompatible with resource-constrained edge environments. This paper presents a systematic design study and performance analysis of FARNet (Facial Attention Residual Network), an efficient deep learning architecture originally proposed for driver facial emotion tracking. Through a structured architectural analysis, this work examines the design rationale behind FARNet's core components: its 2:2:4:2 residual block hierarchy, the ViTinspired large-kernel depthwise convolutions, and the novel Fused Attention (FA) Module employing learnable weighted fusion of channel and spatial attention streams. A comprehensive performance evaluation is conducted across five benchmark datasets — CK+, OuluCASIA, RAF-DB, FER+, and AffectNet — to systematically assess the architecture's accuracy-efficiency trade-off under both lab-controlled and unconstrained in-the-wild conditions. The analysis demonstrates that FARNet achieves perfect classification accuracy of 100% on controlled datasets and a highly competitive 87.65% on the RAF-DB benchmark, utilizing only 3.05 million parameters — a reduction of nearly 74% compared to ResNet-18 baselines. This study provides a critical comparative evaluation against contemporary state-of-the-art models, offering design insights that contribute toward the development of next-generation, parameter-efficient emotion recognition systems for real-time deployment.
Zitieren
APA
Gulia, A., & Dudeja, D. (2026). Design and Performance Analysis of an Efficient Deep Learning Architecture for Facial Emotion Recognition. International Journal of Drug Delivery Technology, 16(40s). https://doi.org/10.25258/ijddt.16.40s.87
Vancouver
Gulia A, Dudeja D. Design and Performance Analysis of an Efficient Deep Learning Architecture for Facial Emotion Recognition. International Journal of Drug Delivery Technology. 2026;16(40s). doi:10.25258/ijddt.16.40s.87
BibTeX
@article{Gulia2026,
author = {Gulia, Abhishek and Dudeja, Deepak},
title = {Design and Performance Analysis of an Efficient Deep Learning Architecture for Facial Emotion Recognition},
journal = {International Journal of Drug Delivery Technology},
year = {2026},
volume = {16},
number = {40s},
doi = {10.25258/ijddt.16.40s.87}
}
EUVIMED