An Optimized Deep Learning Framework for Real-Time Facial Expression Recognition in Neurological Assessment

International Journal of Computer Science and Technology | Vol. 2, No. 9, September 2021 | pp. 65–72

Research Article

Title: An Optimized Deep Learning Framework for Real-Time Facial Expression Recognition in Neurological Assessment

Names of Authors: Elena Rostova¹, Dmitry Ivanov², and Anna Kuznetsova³

Authors’ Affiliations:
¹Department of Computer Science, National Research University Higher School of Economics, Moscow, Russia
²Faculty of Computational Mathematics and Cybernetics, Lomonosov Moscow State University, Moscow, Russia
³Institute of Artificial Intelligence, Saint Petersburg State University, Saint Petersburg, Russia

Abstract: Automated facial expression recognition (FER) systems have emerged as powerful non-invasive tools for monitoring patient neurological responses and cognitive disorders in clinical settings. However, applying FER models to clinical diagnostic workflows requires ultra-low latency processing and robust adaptation to irregular lighting and head pose variations. Standard deep neural networks often suffer from high computational overhead and drop in classification accuracy when deployed on lightweight clinical monitoring units. This paper presents an optimized, real-time FER framework utilizing a modified MobileNetV3 architecture combined with a coordinate attention mechanism. The proposed system retains spatial context by embedding positional info into channel attention, enhancing fine-grained micro-expression extraction. To combat dataset bias, the architecture undergoes transfer learning using a combined dataset of AffectNet and clinical facial sequences. Quantitative evaluations show that the proposed system achieves a classification accuracy of 89.6% across seven core emotional states while restricting model parameter size to 4.2 megabytes (MB). Hardware profiling on an embedded edge device yields an average inference speed of 38.5 frames per second (fps). The framework delivers an efficient, reliable solution for continuous patient monitoring during neurological evaluations without requiring high-end computational infrastructure.

Keywords: Facial expression recognition, Edge intelligence, Deep learning, MobileNet, Coordinate attention, Neurological monitoring

Manuscript Timeline: Received: June 15, 2021; Revised: July 20, 2021; Accepted: August 14, 2021; Published: September 1, 2021

Citation: Rostova, E., Ivanov, D., & Kuznetsova, A. (2021). An optimized deep learning framework for real-time facial expression recognition in neurological assessment. International Journal of Computer Science and Technology, 2(9), 65–72. DOI: 10.46882/2021/IJCST/000021

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read