Abstract
Facial Emotion Recognition (FER) is essential for successful human-computer interaction; however, deploying robust systems on edge devices remains difficult. Recent techniques, such as Vision Transformers (ViTs) and deep ensemble networks, have achieved high accuracy, but suffer from extreme computational overhead and high latency, making them unsuitable for real-time use on limited hardware. The primary challenge lies in maintaining high discriminative power while operating under strict memory and power constraints. To address this, the objective of this research is to develop an efficient Residual Convolutional Neural Network (CNN) optimized for CPU-based inference. The proposed architecture utilizes a hierarchical structure, integrating three consecutive residual blocks with progressively increasing filter depths of 32, 64, and 128. These are engineered to enhance gradient flow and refine feature representation from low-resolution (48 × 48) grayscale images. Comprising only 552,455 parameters and achieving a 12.4 ms latency on standard CPUs, the model balances efficiency and performance. Experimental results on the FER2013 dataset reveal a classification accuracy of approximately 71.4%, outperforming several existing lightweight frameworks. A comprehensive assessment using confusion matrices and ROC curves validates the architecture as a practical solution for real-time affective computing on resource-constrained devices.