Emotion recognition from electroencephalogram signals (EEG-ER) plays an essential role in the development of intelligent human–machine interface applications. This paper proposes an enhancement to the existing EEG_SICNET architecture by integrating the Early-Exit (EE) mechanism, aiming to optimize performance on mobile and resource-constrained devices. The proposed model achieves higher classification accuracy (96.35% compared to 95.24%), while simultaneously reducing the number of parameters by a factor of 1.8 and improving inference speed by approximately 2.5 times on the DEAP dataset. Experimental results on additional benchmark datasets, namely DREAMER and AMIGOS, further validate the consistency of the model’s effectiveness. Compared to recent state-of-the-art deep learning models designed for EEG-ER tasks, the proposed method demonstrates superior performance in terms of both accuracy and inference efficiency, reinforcing its potential for real-time deployment in practical scenarios.