<p>Facial emotion recognition (FER) is an emerging domain in computer vision that enables machines to recognize human facial emotions from expressive features. In recent years, FER has played a significant role in various societal domains such as human–computer interaction, automation systems, education, the automotive industry, and consumer behavior. Although numerous deep learning-based FER models have been developed, many of them face challenges such as complex and heavy architectures, a large number of parameters, and high computational costs. These limitations hinder the practical deployment of existing models in real-world applications. To address these issues, this study proposes a lightweight and efficient deep learning-based model—EfficientNetB3. The proposed model integrates depth-wise separable convolutions, the reverse fusion method (RFM), and a channel attention mechanism to enhance efficiency and robustness, particularly in real-world scenarios. The model’s performance was evaluated on two benchmark datasets: facial emotion recognition-2013 (FER-2013) and Cohn-Kanade+ (CK+), using key metrics such as accuracy, precision, F1-score, root mean squared error (RMSE), and area under the curve (AUC). The optimal results—69.52% accuracy on FER-2013 and 97.10% on CK+—demonstrate a strong balance between efficiency and performance. The model’s practical suitability is further supported by its results on the challenging FER-2013 dataset, along with its optimized model size and inference time.</p>