EmotionLens: Optimizing FER with a Lightweight Residual CNN Architecture

Facial Expression Recognition is a fundamental technology for emotion analysis, mental health diagnosis, and human-computer interaction. Despite notable developments, obtaining accurate emotion classification has remained elusive owing to subtle expression changes and heterogeneous facial characteristics. In this research, a special Residual Convolutional Neural Network model called the EmotionLens, developed specifically for FER without any reliance on pre-trained models, is presented. The model is learned from scratch to categorize eight unique emotions—surprise, fear, neutral, sadness, disgust, contempt, happiness, and anger resulting in increased adaptability and feature specificity. Through extensive experimentation on the AffectNet dataset, a test accuracy of 73.96%, precision of 77.87%, and recall of 69.81% were obtained demonstrating the model's resilience for real-world implementation. These results demonstrate the efficacy of the suggested deep learning method in healthcare, affective computing, and human-centric AI systems. This research helps advance FER by providing a scalable, domain-specific design that achieves both efficiency and interpretability.