Manuscript received July 18, 2025; revised August 5, 2025; accepted September 15, 2025; published January 16, 2026.

Abstract—Online meetings and Virtual Reality (VR) applications require innovative approaches to interpret user emotions and behavior. Since verbal communication is constrained in virtual environments, facial expression analysis is essential for understanding emotional states. Recent research demonstrates that the periocular region provides significant diagnostic information regarding affect and attention, exhibiting pronounced responses to emotional stimuli and offering a more reliable indicator of user state than full-face analysis. This study addresses this gap by evaluating lightweight convolutional neural network architectures—MobileNetV1, MobileNetV2, MobileNetV3, and EfficientNetV2—specifically for periocular-based recognition. Experiments are conducted on the Taiwanese Facial Expression Image Database (TFEID) benchmark, with further validation on the Chinese Face dataset using transfer learning for Android platform deployment. Through a detailed analysis, we evaluate the effectiveness of each architecture based on metrics such as accuracy, precision, recall, and F1-Score, providing insights into their suitability for periocular-based expression recognition. In contrast to earlier studies that employed full-face input, this research proposes a periocular-only approach, rendering it more efficacious in confined environments such as virtual reality or masked-face settings. The findings of this study demonstrate that the MobileNetV3-Small architecture offers an optimal trade-off, attaining an accuracy of 83.62% while sustaining a highly efficient inference time of 16.4 milliseconds per image. Moreover, the deployment of these models on Android devices demonstrates their practicality in real-world settings, particularly in the context of lightweight, mobile-based emotion recognition systems. This research contributes to advancing emotion recognition systems, offering practical and robust solutions for real-world applications.

Keywords—facial expression, periocular area, MobileNet, EfficientNetV2, Taiwanese Facial Expression Image Database (TFEID), Chinese Face dataset

Cite: Sinar B. Ramadhan and David H. Hareva, "Enhancing Facial Expression Recognition: Leveraging MobileNetV3 for Periocular Analysis," Journal of Image and Graphics, Vol. 14, No. 1, pp. 65-75, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.