Real-Time Emotion Recognition Using Lightweight Deep Learning

Abstract

Facial emotion recognition (FER) systems built on deep convolutional neural networks achieve strong accuracy but typically require millions of parameters and substantial compute, making them impractical for deployment on resource-constrained edge devices. This paper proposes MicroExpNet, a lightweight architecture that pairs a MobileNetV3-Small backbone with a channel-wise Squeeze-and-Excitation attention block and a single fully connected classification head. The model is trained through knowledge distillation from a ResNet-50 teacher pretrained on VGGFace2. On the FER2013 benchmark, MicroExpNet reaches 71.2% accuracy with 1.2 million parameters   a 15-fold reduction compared to ResNet-50 (25.6 M) and a 110-fold reduction compared to VGG13 (133 M). Deployed on a Raspberry Pi 4 via ONNX Runtime, the model processes 28 frames per second at 48 × 48 pixel resolution. On AffectNet (8 classes), accuracy is 58.4%. Ablation experiments confirm that knowledge distillation contributes 2.8 percentage points and the attention module adds 1.4 points over the plain MobileNetV3-Small baseline. These results position MicroExpNet as a viable option for real-time affective computing on embedded hardware.