<![CDATA[.Advancements in Artificial Intelligence (AI) and Computer Vision (CV) have driven the development of Human–Computer Interaction (HCI), one of which is Facial Emotion Recognition (FER). However, most FER systems still rely on cloud computing, which poses privacy risks and requires an internet connection. This study focuses on developing an FER system that operates offline on Android devices with high performance and efficient resource usage. The proposed method employs a Convolutional Neural Network (CNN) using two models, a custom CNN and MobileNetV2 based on transfer learning. The FER-2013 dataset was used with three emotion classes: happy, sad, and neutral. Training was conducted using TensorFlow 2.x with the Adam optimizer, a batch size of 32, and 50 epochs. The trained models were then converted into TensorFlow Lite format with INT8 quantization optimization. The MobileNetV2 model achieved an accuracy of 87.72%, with the model size reduced from 18.2 MB to 4.7 MB and inference time under 150 ms per frame. Overall, the combination of MobileNetV2 and TensorFlow Lite proved effective in building a mobile-based emotion recognition system that is fast, efficient, and privacy-preserving through on-device processing. This research can serve as a foundation for further development toward multi-emotion recognition systems and applications supporting mental well-being and adaptive learning in the future]]>
Copyrights © 2026
