TinyFER-Lite: A Small-Data and Lightweight Attention-Based Architecture for Real-Time Facial Emotion Recognition

Abstract

Facial emotion recognition is useful for affective computing, education, human–computer interaction, and assistive systems, but practical models must balance accuracy with model size, latency, and reproducibility. This paper presents TinyFER-Lite, a compact PyTorch pipeline for FER2013-style small and imbalanced facial-expression datasets. The model combines a MobileNetV3-Small feature extractor, Efficient Channel Attention, ImageNet normalization, conservative augmentation, class-weighted focal loss, mixed precision training, and deployment-focused evaluation. On FER2013, the best checkpoint selected by validation macro-F1 obtains 67.07% accuracy, 67.18% macro-F1, and 67.09% weighted-F1 on the held-out PrivateTest split. The trained model has 932,202 trainable parameters, an estimated 3.61 MB size, and measured single-image latency of 8.85 ms on CPU and 7.04 ms on an RTX 3050 GPU in the local evaluation environment. The project does not claim state-of-the-art performance; it provides a reproducible lightweight baseline and a clear foundation for future FER ablations.