<![CDATA[Facial emotion recognition (FER) is a critical component of human-computer interaction, affective computing, and intelligent surveillance systems. Existing deep learning approaches, while achieving high accuracy, are often computationally expensive and unsuitable for deployment on resource-constrained or real-time systems. In this paper, we present LightEmoNet, a lightweight Convolutional Neural Network (CNN) architecture specifically designed for efficient and accurate facial emotion recognition. Our model is trained on the FER2013 benchmark dataset, which contains 35,887 grayscale images distributed across seven emotion classes: Happy, Neutral, Sad, Fear, Angry, Surprise, and Disgust. To address the inherent class imbalance within the dataset, we employ a dual strategy combining class-weighted loss penalization with targeted data augmentation applied selectively to underrepresented categories. The proposed architecture totals approximately 2.1 million trainable parameters and occupies only 8.3 MB on disk, making it deployable on edge and embedded platforms without GPU acceleration. Experimental results demonstrate that LightEmoNet achieves a training accuracy of 91.0% and a validation accuracy of 88.5% on the FER2013 test split, with an average inference latency of 4.2 ms per image on a standard CPU. The model exhibits robust performance across all seven emotion classes while maintaining a compact footprint suitable for real-time inference. These findings confirm that lightweight CNNs, when paired with principled augmentation strategies, can achieve competitive performance without the overhead of large-scale deep models.]]>
