The search for efficient deep learning architectures for emotion recognition using EEG signals has drawn great interest due to applications in healthcare, education, and intelligent interaction. These models must meet three key requirements: achieving high accuracy with fewer electrodes (32, 14, or even 5), maintaining stable performance across frequency bands, and being lightweight enough for deployment on low-resource devices. This paper proposes EEG_SICNET, an enhanced 1D-CNN integrated with Squeeze and Excitation and Inception blocks to optimize EEG signal processing. Experiments on DEAP, DREAMER, and AMIGOS datasets demonstrate EEG_SICNET's compact size (40.14 MB), stable performance across frequency bands, and accuracy up to 83% with 5 electrodes. Additionally, it achieves over 72% accuracy when deployed on a Raspberry Pi 4 with 14-channel input, outperforming recent methods on the DEAP dataset.