ARECA-Lite: A lightweight modified ArecaNet with reduced complexity for real-time robust facial emotion recognition

Facial emotion recognition (FER) in e-learning environments faces a persistent tension between recognition performance and deployment efficiency; high-performing architectures typically impose prohibitive computational costs that preclude real-time use on commodity hardware. This paper presents ARECA-Lite, a lightweight variant of ArecaNet designed to resolve this tension in resource-constrained deployments. ARECA-Lite integrates a truncated MobileNetV3-Small backbone with dual VGGFace2-pretrained InceptionResNetV1 sub-branches, a wavelet pooling module for structured frequency-domain decomposition, and a modified Assembled Residual Enhanced Cross-Attention (A.R.E.C.A.) module composed of two parallel RECA blocks. A FACS-guided offline preprocessing pipeline, targeted disgust oversampling, and a composite Dice-BCE loss collectively address class imbalance and low-contrast action unit (AU) evidence. Experiments are conducted on FER2013. ARECA-Lite achieves 82.41% accuracy and an 82.50% macro F1-score, outperforming a retrained ArecaNet baseline under the same experimental conditions by 7.10 and 9.43 percentage points, respectively, while reducing parameter count from 24.95 M to 4.69 M, GPU inference latency from 21.83 ms to 5.44 ms, and computational cost from 6.57 GFLOPs to 1.64 GFLOPs. The preprocessing and augmentation pipeline adds 14.88 percentage points in accuracy relative to raw-data training. Evaluation is limited to FER2013; broader validation is left to future work. These results demonstrate that competitive, class-balanced facial emotion recognition is achievable without the substantial computational overhead of larger attention-based models, making ARECA-Lite well-suited to low-latency affective computing on commodity hardware.